Integrate the latest version of CNMEM

This version comes with two fixes.

1/ Fix a warning due to the visibility attribute needed by Theano.
We changed const char* CNMEM_API to const char CNMEM_API* in the
declaration of cnmemGetErrorString to have the attribute between
char and star. It fixes the warning.

2/ Fix a compilation issue with CUDA 5.0 due to the missing
cudaStreamGetFlags function. It is #defined out when the Runtime
is less than 5.5 (5050 version).

Flake8 misc

changing sort.grad to work with other ndim

Mixed stuff

Speed up Scan optimizations

[BUG, TEST] Fix Sng handling of nan and fix related test

Fix compilation on windows with VS 2012

Don't apply alpha_merge and output_merge when the proc node has more than one client.

ENH: allow numpy.*(theano_var) to build a graph

Allow slicing of tensor variables to return symbolic shape refs

pep8 fix in master

Flake8 gof sequel

Further work on issue #3018.

Revert the change that would broadcast the output if necessary and just don't apply the opt in that case.

than one client.  Also apply output_merge in the case when the new
output has to be broadcasted.

Make travis fail when the first part fail. Not sure if this kill/don't start not started part. But should tell the user more rapidly in github interface