Bouchnic truedot rebase

Add documentaiton about the MPI and load ops

Fix test that fail depending of the seed

Use dtype specific comparison

This should remove dtype specific to float32 in the nightly buildbot.

print more detailed error message to help fix the test error related to ...

WIP:Speeding up GpuAdvancedIncSubTensor1 by writing fast Cuda Code

Add utt.assert_allclose that raise a more informative error.

This is needed in the Scan tests.

The numpy.allclose() use a tolerate too high for float32.

This will help remove an error that appear with different seed in the tests.

Fix tests

Rebase gh-1368

fix for variable size array problem in Windows compiler

[WIP] Use the new grad interface.

This print the time used for each batch.

The old range was causing nan in float32.

When the output was of size 0, we where returning a numpy ndarray
instead of a CudaNdarray.

No news, as this is a fix to something new since the last release.

Make GpuCrossentropySoftmaxArgmax1HotWithBias and GpuCrossentropySoftmax1HotWithBiasDx work for bigger inputs.

I don't know if argsort should have a grad defined, so I mark it as not implemented.