Closes ticket #518.

make the split of too huge GpuElemwise check about the number of dimensions of the inputs. test this too.

Added fix suggested by Pascal for the tests when dealing with the exception thrown by Subtensor (e_invalid)

added test for same bug

more helpful.

[optimization][dot/gemv -> gpu_dot22/gpu_gemm] Two new optimizations that will convert a gemv or dot ( i.e. vector - matrix
dot products) to GpuGemm and GpuDot22 by reshaping the vector to matrix

If it isn't raise warning and default to py-linkear.

Remove some scipy dependency. We disable the perform method of the convolution when scipy is not available.