We compute A B -> C with matrices A, B, C

With the blas (cblas) interface it supports matrices that adhere to
certain criteria. They should be contiguous on one dimension (stride=1).

We glance a little at how numpy does this to try to catch all cases.

In short, we accept A, B contiguous on either axis (row or column
major). We use the case where C is (weakly) row major, but if it is
column major we transpose A, B, C => A^t, B^t, C^t so that we are back
to the C row major case.

(Weakly = contiguous with stride=1 on that inner dimension, but stride
for the other dimension can be larger; to differentiate from strictly
whole array contiguous.)

Minor change to the gemv function, no functional change, only updating
due to the refactoring of blas layout functions.

Fixes #1278

If we have a matrix of dimension say 5 x 5, BLAS requires the leading
stride to be >= 5. Smaller cases are possible for read-only array views
in ndarray(broadcasting and custom strides).

In this case we mark the array as not BLAS compatible

Using cblas we can simplify this further to a more satisfying
translation (from ndarray to BLAS), much simpler logic.

Avoids creating and handling an extra layer of array views.

Add a crate with a mock blas implementation, so that we can assert that
cblas_sgemm etc are called (depending on memory layout).