Remove uneeded import from code example

target of :paramref:`.relationship.secondary` for the purposes
of creating very complex :func:`.relationship` join conditions.
The change includes adjustments to query joining, joined eager loading
to not render a SELECT subquery, changes to lazy loading such that
the "secondary" target is properly included in the SELECT, and
changes to declarative to better support specification of a
join() object with classes as targets.

This had me reread the code twice to see if I missed why the import was
present.

- tests
- full documentation, changelog, new in 0.9 announcement

against the remaining tables within the given metadata.

:class:`.DeferredReflection`
to define classes pending for reflection, if some subset of those
classes were discarded before the :meth:`.DeferredReflection.prepare`
method were called to reflect and map the class, a strong reference
to the class would remain held within the declarative internals.
This internal collection of "classes to map" now uses weak
references against the classes themselves.

attribute on declarative to simply refer directly to an :class:`.InstrumentedAttribute`
on a superclass or on the class itself, and it
acts more or less like a synonym; in 0.9, this fails to set up enough
bookkeeping to keep up with the more liberalized backref logic
from :ticket:`2789`.  Even though this use case was never directly
considered, it is now detected by declarative at the "setattr()" level
as well as when setting up a subclass, and the mirrored/renamed attribute
is now set up as a :func:`.synonym` instead. [ticket:2900]

:class:`.Column` is mapped multiple times under different properties
(which typically should be a :func:`.synonym` instead) or if two
or more :class:`.Column` objects are given the same name, raising
a warning if this condition is detected. [ticket:2828]