Added GBMV with unit test, fixed GEMV function

[tab58]

PTAL @rreusser.

[rreusser]

Awesome! Closely related things that may be useful for testing, comparison, reference, etc:

• ndarray-diagonal
• ndarray-band
• solve-banded
• solve-tridiagonal
• solve-periodic-tridiagonal

The banded stuff is pretty straightforward, but gets to more of those details where strided vs. getter/setter vs. vanilla array vs. ndarray stuff starts to surface.

But all of that aside, specific thoughts:

• As you've noticed, ndarrays mostly alleviate the need for transposes since, roughly speaking, gbmv(..., A.transpose(1, 0), ...) is way easier than coding gbmv to account for transposes. So lots of the upper/lower flags of blas may just not be relevant.
• Actually, that's most of it. Let's keep these getter/setter based, test, publish, benchmark, and optimize by the same pattern.

:+1:

[rreusser]

Also, need to investigate exactly how BLAS handles compressed formats like diagonal matrices. Does BLAS deal strictly in dense full-storage formats and things like compressed banded storage are a completely separate problem?

[tab58]

AFAIK, BLAS only handles full matrices, with the exception of triangular packed matrices. There may be support in other libraries like LAPACK to take advantage of compressed banded matrices, but I think that's outside the scope of this library.

Are there any other things that I need to do on this before this can get merged in?