See README.unittests for information on how to run
the tests.  [ticket:970]

fail when default-holding columns are flushed.

      in conjunction with large mapper graphs, large numbers of
      objects:

      - The Session's "weak referencing" behavior is now *full* -
        no strong references whatsoever are made to a mapped object
        or related items/collections in its __dict__.  Backrefs and
        other cycles in objects no longer affect the Session's ability
        to lose all references to unmodified objects.  Objects with
        pending changes still are maintained strongly until flush.
        [ticket:1398]

        The implementation also improves performance by moving
        the "resurrection" process of garbage collected items
        to only be relevant for mappings that map "mutable"
        attributes (i.e. PickleType, composite attrs).  This removes
        overhead from the gc process and simplifies internal
        behavior.

        If a "mutable" attribute change is the sole change on an object
        which is then dereferenced, the mapper will not have access to
        other attribute state when the UPDATE is issued.  This may present
        itself differently to some MapperExtensions.

        The change also affects the internal attribute API, but not
        the AttributeExtension interface nor any of the publically
        documented attribute functions.

      - The unit of work no longer genererates a graph of "dependency"
        processors for the full graph of mappers during flush(), instead
        creating such processors only for those mappers which represent
        objects with pending changes.  This saves a tremendous number
        of method calls in the context of a large interconnected
        graph of mappers.

      - Cached a wasteful "table sort" operation that previously
        occured multiple times per flush, also removing significant
        method call count from flush().

      - Other redundant behaviors have been simplified in
        mapper._save_obj().

      not mapping an attribute that already had class-level
      name present.

of the same direction, i.e. ONETOMANY or MANYTOONE,
is now detected, and an error message is raised.
Saves crazy CircularDependencyErrors later on.

instrumented descriptor is passed, raises
"Invalid column expession".

fully establish instrumentation for subclasses where the mapper
was created after the superclass had already been fully
instrumented. [ticket:1292]

Session methods have been deprecated, replaced by
"expunge_all()" and "add()".  "expunge_all()" has also
been added to ScopedSession.

- Added a new `relation()` keyword `back_populates`.  This allows configuation of backreferences using explicit relations. [ticket:781]  This is required when creating bidirectional relations between a hierarchy of concrete mappers and another class. [ticket:1237]
- Test coverage added for `relation()` objects specified on concrete mappers. [ticket:1237]
- A short documentation example added for bidirectional relations specified on concrete mappers. [ticket:1237]
- Mappers now instrument class attributes upon construction with the final InstrumentedAttribute object which remains persistent.  The `_CompileOnAttr`/`__getattribute__()` methodology has been removed.  The net effect is that Column-based mapped class attributes can now be used fully at the class level without invoking a mapper compilation operation, greatly simplifying typical usage patterns within declarative. [ticket:1269]
- Index now accepts column-oriented InstrumentedAttributes (i.e. column-based mapped class attributes) as column arguments.  [ticket:1214]
- Broke up attributes.register_attribute into two separate functions register_descriptor and register_attribute_impl.    The first assembles an InstrumentedAttribute or Proxy descriptor, the second assembles the AttributeImpl inside the InstrumentedAttribute.  register_attribute remains for outside compatibility.  The argument lists have been simplified.
- Removed class_manager argument from all but MutableScalarAttributeImpl (the branch had removed class_ as well but this has been reverted locally to support the serializer extension).
- Mapper's previous construction of _CompileOnAttr now moves to a new MapperProperty.instrument_class() method which is called on all MapperProperty objects at the moment the mapper receives them. All MapperProperty objects now call attributes.register_descriptor within that method to assemble an InstrumentedAttribute object directly.
- InstrumentedAttribute now receives the "property" attribute from the given PropComparator.  The guesswork within the constructor is removed, and allows "property" to serve as a mapper compilation trigger.
- RelationProperty.Comparator now triggers compilation of its parent mapper within a util.memoized_property accessor for the "property" attribute, which is used instead of "prop" (we can probably remove "prop").
- ColumnProperty and similar handle most of their initialization in their __init__ method since they must function fully at the class level before mappers are compiled.
- SynonymProperty and ComparableProperty move their class instrumentation logic to the new instrument_class() method.
- LoaderStrategy objects now add their state to existing InstrumentedAttributes using attributes.register_attribute_impl.  Both column and relation-based loaders instrument in the same way now, with a unique InstrumentedAttribute *and* a unique AttributeImpl for each class in the hierarchy.  attribute.parententity should now be correct in all cases.
- Removed unitofwork.register_attribute, and simpified the _register_attribute methods into a single function in strategies.py.  unitofwork exports the UOWEventHandler extension directly.
- To accomodate the multiple AttributeImpls across a class hierarchy, the sethasparent() method now uses an optional "parent_token" attribute to identify the "parent".  AbstractRelationLoader sends the MapperProperty along to serve as this token.  If the token isn't present (which is only the case in the attributes unit tests), the AttributeImpl is used instead, which is essentially the same as the old behavior.
- Added new ConcreteInheritedProperty MapperProperty.  This is invoked for concrete mappers within _adapt_inherited_property() to accomodate concrete mappers which inherit unhandled attributes from the base class, and basically raises an exception upon access.  [ticket:1237]
- attributes.register_attribute and register_descriptor will now re-instrument an attribute unconditionally without checking for a previous attribute.  Not sure if this is controversial. It's needed so that ConcreteInheritedProperty instrumentation can be overridden by an incoming legit MapperProperty without any complexity.
- Added new UninstrumentedColumnLoader LoaderStrategy.  This is used by the polymorphic_on argument when the given column is not represented within the mapped selectable, as is typical with a concrete scenario which maps to a polymorphic union.  It does not configure class instrumentation, keeping polymorphic_on from getting caught up in the new concrete attribute-checking logic.
- RelationProperty now records its "backref" attributes using a set assigned to `_reverse_property` instead of a scalar.  The `back_populates` keyword allows any number of properties to be involved in a single bidirectional relation.  Changes were needed to RelationProperty.merge(), DependencyProcessor to accomodate for the new multiple nature of this attribute.
- Generalized the methodology used by ManyToManyDP to check for "did the other dependency already handle this direction", building on the `_reverse_property` collection.
- post_update logic within dependency.py moves to use the same methodology as ManyToManyDP so that "did the other dependency do this already" checks are made to be specific to the two dependent instances.
- Caught that RelationProperty.merge() was writing to instance.__dict__ directly (!) - repaired to talk to instance_state.dict.
- Removed needless eager loading example from concrete mapper docs.
- Added test for [ticket:965].
- Added the usual Node class/nodes table to orm/_fixtures.py, but haven't used it for anything yet.   We can potentially update test/orm/query.py to use this fixture.
- Other test/documentation cleanup.