The "raise_on_sql" option differentiates from "raise" in that
firing a lazy loader is OK as long as it does a simple
get from identity map.   Whereas "raise" is more useful
for the case that objects are to be detached.

As part of this, refactors the strategy initiation logic
a bit so that a LoaderStrategy itself knows what "key" was used
to create it, thus allowing variants of a single strategy
based on what the "lazy" argument is.  To achieve this we
have to also get rid of _get_strategy_by_cls().

Everything here is internal with the one exception of an apparently
undocumented, but not underscored, "strategy_class" key
on relationship().   Though it's not clear what
"strategy_class" accomplishes; at this point the strategy
system is extensible using Property.strategy_for().

Fixes: #3812
Change-Id: I812ad878ea5cf764e15f6f71cb39eee78a645d88

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

any speed improvements :(.   code is in a much better place to be run into
C, however
- The ``proc()`` callable passed to the ``create_row_processor()``
method of custom :class:`.Bundle` classes now accepts only a single
"row" argument.
- Deprecated event hooks removed:  ``populate_instance``,
``create_instance``, ``translate_row``, ``append_result``
- the getter() idea is somewhat restored; see ref #3175

:class:`.SynonymProperty` and :class:`.ComparableProperty`.
- The ``info`` parameter has been added as a constructor argument
to all schema constructs including :class:`.MetaData`,
:class:`.Index`, :class:`.ForeignKey`, :class:`.ForeignKeyConstraint`,
:class:`.UniqueConstraint`, :class:`.PrimaryKeyConstraint`,
:class:`.CheckConstraint`.

fixes #2963

sqlalchemy/orm, sqlalchemy/event, sqlalchemy/testing

to get all flake8 passing

calling ``MyClass.attribute`` will not require that the configure
mappers step has occurred, e.g. it will just work without throwing
any error. [ticket:2935]