fix context problem with pycuda by checking it is imported before.

Opt

We already checked that the node is still in the env before we process it.

Also add the pycuda.init flag to force its import.

It is not needed to set it to True in the buildbot, as all test file
get imported before test start to run. So this init PyCUDA first if
the theano flag device and init_gpu_device are not used.

Bug reported by Nicolas regarding taps of scan sequences

ajout parametre keepdims

Small

Time opt

The bug was reported by Nicolas, and it happens when you have a single
tap for a sequence that is different from 0.

Accept profile=False and when we profile, preload the c module cache to don't have it include in optimization time.

fix for the bug involving multiple subtensor on a scan.

The consequence was warning printed to the user and scan use more memory.

Revert the grad of Mul. The new version fixed a probably not happening c...

I added assert to have an error for the case not implemented.

adding support for numpy.newaxis in basic indexing