multiple data type support for GPUMultinomialFromUniform output

Improve the example for strict=True in the scan() doc part 2.

Fix test for Python 3 on 32-bit

Interface addition and cleanup

replace multinomial_wo_replacement by choice

Change MRG choice default of a to be None. We can do this as in the old back-end, when a wasn't provided and pvals was, a=2, but it was only valid when that match pvals shapes. So it was duplicate information. As MRG have pvals mandatory, to work it need a matching value for a.

Merge optimizer 5319

Add C code for CPU and GPU for tensor.Split op.

Fix gpu integers in scan.

validate numpy1.12 compatibility changes

casting to int could lead to issues in the following case:
- 64-bit Windows
- python 2
- shape bigger than 2**31

In that case, int would be 32-bit, but shapes could be bigger and stored
in long.

Fix CAReduce.c_code_cache_version_apply
Calling `append` on a tuple raised AttributeError, which was mistaken
for a missing `c_code...` function altogether, and the version was
ignored.

Done in both CPU and GPU code.

Speed up one scan opt

Mixed

Typo in sharedvalue.py

* Added missing softsign import

* Update __init__.py