Both BLAS and Numpy were trying to be too clever, BLAS by wanting a
pointer to the beginning of the memory chunk (even if it will start from
the end, and Numpy by providing a pointer to the first element, that is
at the end of the memory buffer.

PyArray_GETCONTINUOUS considers 0-sized arrays as contiguous,
even if they have negative strides on the 0-length dimension for
instance, which confuses BLAS.

And it seems a good idea, since several are failing...

It should now support matrices with negative strides, and non-unit
strides in general, by making a copy in the appropriate cases.

Also allow f-contiguous mat in Gpu sger.

Instead of just crashing the program.

This enables calling CudaNdarray_Copy on a const CudaNdarray*,
which was previously impossible.

It is an error in recent versions of Vim.

Make a machanism to don't have Alloc and GpuAlloc constant folded in som...

pep8 fix + added verbose exception

Small fixes

Add shape inference for sparse/basic.py Ops

A few PEP8 fixes, string formatting and cosmetic stuff.