This was done in order to greatly simplify a new optimization to serialize
multiple IncSubtensors of a common base.  This is a case that happens if you
compute more than one subtensor from a single variable and then take a gradient.
Serialization helps because then all the IncSubtensors can be done inplace.

SetSubtensor.

1) Fixed __eq__ to work when other is not of same type as self.
2) Optimized TensorConstantSignature to compare data.sum() before comparing all elements.

different ids that compare equal.

[recursive] handling of code versioning for ops like Elemwise, CAReduce, etc.

Changed the default cache version of Op to ().

Added c_code_cache_version() functions to many Ops that were using the default
before.

scalar/basic.  There is also a cast function, which is used in the grad() of ops
that might upcast their arguments, to downcast the corresponding gradients.

types.