size of the many per-column objects we're hitting, but somehow the overall memory is
hardly being reduced at all in initial testing

through all possible set values

assume that the empty string, or a set with a single empty string
value, is in fact a set with a single empty string; instead, this
is by default treated as the empty set.  In order to handle persistence
of a :class:`.mysql.SET` that actually wants to include the blank
value ``''`` as a legitimate value, a new bitwise operational mode
is added which is enabled by the
:paramref:`.mysql.SET.retrieve_as_bitwise` flag, which will persist
and retrieve values unambiguously using their bitflag positioning.
Storage and retrieval of unicode values for driver configurations
that aren't converting unicode natively is also repaired.
fixes #3283

isinstance() calls, express intent more clearly

- test_schema_2 is only on PG and doesn't need a drop all, omit this for now
- py3k has exception.args[0], not message

sort_tables_and_constraints function.
- The DDL generation system of :meth:`.MetaData.create_all`
and :meth:`.Metadata.drop_all` has been enhanced to in most
cases automatically handle the case of mutually dependent
foreign key constraints; the need for the
:paramref:`.ForeignKeyConstraint.use_alter` flag is greatly
reduced.  The system also works for constraints which aren't given
a name up front; only in the case of DROP is a name required for
at least one of the constraints involved in the cycle.
fixes #3282

to complement the :attr:`.Table.foreign_keys` collection,
as well as :attr:`.ForeignKeyConstraint.referred_table`.

bump

foo