- flush() refactor merged from uow_nontree branch r3871-r3885

- flush() refactor merged from uow_nontree branch r3871-r3885
- topological.py cleaned up, presents three public facing functions which
return list/tuple based structures, without exposing any internals.  only
the third function returns the "hierarchical" structure.  when results
include "cycles" or "child" items, 2- or 3- tuples are used to represent
results.
- unitofwork uses InstanceState almost exclusively now.  new and deleted lists
are now dicts which ref the actual object to provide a strong ref for the
duration that they're in those lists.  IdentitySet is only used for the public
facing versions of "new" and "deleted".
- unitofwork topological sort no longer uses the "hierarchical" version of the sort
for the base sort, only for the "per-object" secondary sort where it still
helps to group non-dependent operations together and provides expected insert
order.  the default sort deals with UOWTasks in a straight list and is greatly
simplified.  Tests all pass but need to see if svilen's stuff still works,
one block of code in _sort_cyclical_dependencies() seems to not be needed anywhere
but i definitely put it there for a reason at some point; if not hopefully we
can derive more test coverage from that.
- the UOWEventHandler is only applied to object-storing attributes, not
scalar (i.e. column-based) ones.  cuts out a ton of overhead when setting
non-object based attributes.
- InstanceState also used throughout the flush process, i.e. dependency.py,
mapper.save_obj()/delete_obj(), sync.execute() all expect InstanceState objects
in most cases now.
- mapper/property cascade_iterator() takes InstanceState as its argument,
but still returns lists of object instances so that they are not dereferenced.
- a few tricks needed when dealing with InstanceState, i.e. when loading a list
of items that are possibly fresh from the DB, you *have* to get the actual objects
into a strong-referencing datastructure else they fall out of scope immediately.
dependency.py caches lists of dependent objects which it loads now (i.e. history
collections).
- AttributeHistory is gone, replaced by a function that returns a 3-tuple of
added, unchanged, deleted.  these collections still reference the object
instances directly for the strong-referencing reasons mentiontioned, but
it uses less IdentitySet logic to generate.