intelligible, given the fixes for ref #3623.  unfortunately the system
is still quite weird even though it was rewritten to be... less weird

system failed to accommodate for multiple :func:`.undefer_group`
loader options in a single query.   Multiple :func:`.undefer_group`
options will now be taken into account even against the same
entity. fixes #3623

the check added for a polymorphic joinedload from a
poly_subclass->class->poly_baseclass connection would fail for the
scenario of class->poly_subclass->class.
fixes #3611

would not bump a version id counter when in use.  The experience
here is still a little rough as the original version id is required
in the given dictionaries and there's not clean error reporting
on that yet.
fixes #3610

insert-default holding columns not otherwise included in the SET
clause (such as primary key cols) to get rendered into the RETURNING
even though this is an UPDATE.

- Major fixes to the :paramref:`.Mapper.eager_defaults` flag, this
flag would not be honored correctly in the case that multiple
UPDATE statements were to be emitted, either as part of a flush
or a bulk update operation.  Additionally, RETURNING
would be emitted unnecessarily within update statements.

fixes #3609

cause a subsequent call to the :meth:`.Query.with_parent` method to
fail. fixes #3606
- add mark-as-fail test for #3607

primary key before emitting an INSERT, and merges distinct objects with
duplicate primary keys together as they are encountered, which is
essentially semi-deterministic at best.   This behavior
matches what happens already with persistent objects.
fixes #3601

fail to emit an UPDATE in the case where the attribute were set to
None and not previously loaded.
fixes #3599