a much more modest outcome than what we started with.   The
work of create_row_processor() for ColumnProperty objects
is essentially done at query setup time combined with some
lookups in _instance_processor().
- to allow this change for deferred columns, deferred columns
no longer search for themselves in the result.   If they've been
set up as deferred without any explicit directive to undefer them,
then this is what was asked for.  if we don't do this,
then we're stuck with this performance penalty for all deferred
columns which in the vast majority of typical use cases (e.g. loading
large, legacy tables or tables with many/large very seldom
used values) won't be present in the result and won't be accessed at all.

the :func:`.orm.load_only` option to cover all attributes not
explicitly mentioned, now takes into account the superclasses
of a given entity, if that entity is mapped with inheritance mapping,
so that attribute names within the superclasses are also omitted
from the load.  Additionally, the polymorphic discriminator column
is unconditionally included in the list, just in the same way that
primary key columns are, so that even with load_only() set up,
polymorphic loading of subtypes continues to function correctly.
fixes #3287

for memoization on a class that uses slots.
- apply many more __slots__.  mem use for nova now at 46% savings

allowing us to move to __slots__

non-recommended use case of replacing a relationship on a fixed
mapped class many times, referring to an arbitrarily growing number of
target mappers.  A warning is emitted when the old relationship is
replaced, however if the mapping were already used for querying, the
old relationship would still be referenced within some registries.
fixes #3251