disable complex support in gpu reduce.

Started it, but I didn't finished it and I have other stuff to finish...

[ENH] speed up reduction on the colums of matries.

Add function_dump() to help user give us reproducable example.

Add check on the output of NumPy.

Add warning if the cuda lib dir is in the -L path.

[BUG] Do a copy when NumPy didn't allocate new memory.

An opt will remove this case if the repeat pattern is a constant 1.

Modifications in order to pass the tests in pylearn2/cross_validation/tests/test_cross_validation.py

Stop adding -L/usr/local/cuda/lib{,64} as argument to nvcc.

Switch to cublas v2

This was never needed and also causes the errors with fatbinwrap_65.

Opt that remove useless tile

Speed up Debugmode