simple but unusual system allows for a dramatic savings in Python
overhead for the construction and processing of orm :class:`.Query`
objects, from query construction up through rendering of a string
SQL statement.
fixes #3054

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.

this would cause LazyLoader to be initialized twice per property.

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

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

- prop.compare() isn't needed; replace with prop._with_parent()
for relationships
- update docs in orm/interfaces

instances(), using a named tuple it can assign to directly.  this way
we never have to worry about that structure changing anymore, though
we are still having it append (key, fn) which is kind of awkward.
- inline _populators() into instance(), it's a little verbose but
saves an fn call