Vectorise matrix similarity functions

[tornikeo]

Addressing #637. I rewrote the njitted functions using plain matmul operations, using numpy. This is around x50 times faster (2s for 1k pairwise 1k peaks, vs 40ms).

[niekdejonge]

@florian-huber Seems like a great change to me, but I am less familiar with this part of the code base. Could you check and merge this PR?

[niekdejonge]

@tornikeo please feel free to add yourself to the citation.cff file, btw!

[tornikeo]

Oh, wait a sec... Once I remove pytorch matmul from the picture, numpy seems only around 2x faster, not 50x times faster. This is all on CPU. It's weird to have such a difference between cpu-pytorch and numpy.

Still investigating this. Just a 2x acceleration might not be worth more obscure code.

[niekdejonge]

Hmm yes, surprising that the difference is that large. Still 2x acceleration can be worth it, I will leave it up to @florian-huber. Let us know what comes out of the investigation :)

[tornikeo]

Found the culprit. It's because numpy only uses BLAS for floaty types. There's no integer support for BLAS matmul. Still numpy is slower than torch. With this, we arrive to around 10x speedup.

I've also expanded the comparison notebook to compare all three approaches (numpy, torch and njit).