Fixes a performance digression caused by Blas Ops that did not leverage C-thunking.

The CLinkerOp and CLinkerObject like types were moved into link.c.interface to get them out of gof.
PureOp and Op were merged (as Op) which means that some C-related methods are now declared on all Ops.

+ BITWIDTHs are now constants in utils.
+ Lambdas and local functions for default/filter/validate are now module-level local functions. This fixes pickleability (closes #240).

Move linkers to link module

The `CVM` class definition was also split off into `theano.link.c.cvm`.

Closes #223

Also adds some type hints.

The comment said to move Container to compile, but it is directly and
exclusively used by the Linkers.  Some type hints were added and Container now
requires kwarg specification for the optional arguments.

Admittedly, the `undef = object()` is a cool trick, but `allow_gc` is simply a
null-able boolean.  Also, the `make_function` method can probably be removed, so
I marked it as deprecated.

Clean config importing

This untangles the `theano.configdefaults` module from the `theano.gof`
sub-package, thereby removing an import cycle.

Closes #94

These helper functions where in `theano.misc.windows`, but they're not actually
specific to Windows and they're clearly generic "utils" style functions.

This way all sub-packages get the chance to register their config variables
without needing to be imported in theano.configdefaults.

* Move JAX linker to new "link" module
closes #188

* Move JAX tests to link module and update import path

* Remove `int_division` from `configdefaults.py`

* * remove `div_proxy`, replace by either `true_div` or `int_div`

* TestArithmeticCast :  remove references to `config.int_division` and interger division

* Remove __intdiv__ method from scalar.basic.py

* Rebase

* Fix merging conflicts

Add new RandomVariable Op and optimizations

This optimization does *not* preserve equality between the numeric
results of the untransformed and transformed graphs when the RNGs and seeds are
equal.  The reason is that the underlying sampler methods themselves are not
implemented in Theano, so we cannot apply the requisite DimShuffle-like
operations to the intermediate samples used to generate multiple replications
and/or independent variates.

For example, sampling a normal of size (3, 2) requires a draw of size (3, 2)
from a standard normal and we can't transpose that (3, 2) array.  If we could,
then we would be able to maintain numerical equality between graphs.