Add cdef nogil sgemm and saxpy functions

[danieldk]

The motivation for adding these functions is to be able to use them from cdef nogil code from different BLAS implementations interchangeably. The signature mirrors CBLAS as closely as possible, with the following simplifications to sgemm:

• Remove the order argument, always assume C-order.
• Since a conjugate transpose is not possible on a float matrix, specify transpose using booleans.

We also use 32-bit integer sizes and increments as the lowest-common denominator of different BLAS versions.

[shadeMe]

LGTM.

Just one question: Would it make sense to also include variants of the two functions that take double precision arguments, for the sake of completeness? The original gemm and axpy functions support both.

[danieldk]

Just one question: Would it make sense to also include variants of the two functions that take double precision arguments, for the sake of completeness? The original gemm and axpy functions support both.

We currently don't use them in spaCy, but I think it makes sense to add these for completeness. So I'll add them.

[danieldk]

We currently don't use them in spaCy, but I think it makes sense to add these for completeness. So I'll add them.

Added.