-> crossentropy optimization still works when storing labels on the GPU.

gpuelemwise

Get rid of references to the older THEANO_* environment variables in the code comments and inline doc.

Change type of config.unittests.rseed so it accepts the value "random", as THEANO_UNITTEST_SEED did.

* removed numpy as N, tensor as T
* wrapped docstrings to 80cols
* moved getFilterOutShp to staticmethod
* removed some useless code (old convolve2 function)

Reimplemented unittests for conv.

Speed tests will go in separate file (to come)

-side effect, the cuda test are not enabled by default, will be on the build bot.
-also, now the use fct don't have a default value.

fixes to make the behaviour of the cuda shared constructor more strict on accepting only float32 ndarrays

lower profiler overhead by moving computation into the printing phase instead of the recording phase.