primary key change within a savepoint block would not
participate in being restored to their former state (not in
session, in session, previous PK) after the outer transaction
were rolled back. fixes #3108

as the attribute and probably just replaced itself, so that is
now _set_enable_single_crit
- as a side effect of the main issue fixed here, correct the case in
adjust_for_single_inheritance where the same mapper appears more
than once in mapper_adapter_map; run through a set() for uniqueness.
- Fixed bug in subquery eager loading in conjunction with
:func:`.with_polymorphic`, the targeting of entities and columns
in the subquery load has been made more accurate with respect
to this type of entity and others. Fixes #3106

of :ticket:`3060` from verision 0.9.5.  A self-referential relationship
with lazy='dynamic' would raise a TypeError within a flush operation.
fixes #3099

when we do a get; we return the None as always but we leave the dict blank
and the loader callable still in place.  The case for this implicit get on a pending object is not
super common and there really should be no change in state at all when this
operation proceeds.   This change is more dramatic as it reverses
a behavior SQLA has had since the first release.
fixes #3061

mapper is implicitly combining one of its column-based attributes
with that of the parent, where those columns normally don't necessarily
share the same value.  This is an extension of an existing check that
was added via :ticket:`1892`; however this new check emits only a
warning, instead of an exception, to allow for applications that may
be relying upon the existing behavior.
fixes #3042

columns are always added to the list of columns to be "undeferred";
otherwise, the ORM can't load the row's identity.   Apparently,
one can defer the mapped primary keys and the ORM will fail, that
hasn't been changed.  But as load_only is essentially saying
"defer all but X", it's more critical that PK cols not be part of this
deferral. fixes #3080

of work such that loading for related many-to-one objects is slightly
more aggressive, in the case of a graph of self-referential objects
that are to be deleted; the load of related objects is to help
determine the correct order for deletion if passive_deletes is
not set.
- revert the changes to test_delete_unloaded_m2o, these deletes do in fact
need to occur in the order of the two child objects first.

implicitly initialized to None via first access; this action,
which has always resulted in a population of the attribute,
now emits an attribute event just like any other attribute set
operation and generates the same kind of history as one.  Additionally,
many mapper internal operations will no longer implicitly generate
these "None" values when various never-set attributes are checked.
These are subtle behavioral fixes to attribute mechanics which provide
a better solution to the problem of :ticket:`3060`, which also
involves recognition of attributes explicitly set to ``None``
vs. attributes that were never set.
fixes #3061