Speed up cross-correlation with FFT convolution

[comavius]

Overview

This PR replaces the O((n1 + n2) * nframes * numcep) sequential offset scan with O((n1 log n1 + n2 log n2) * numcep) FFT convolution.

Benchmarking

I compared the execution time of find_offset_between_buffers between this branch and the main branch.

The benchmark uses the full 1812 Overture audio as the reference buffer and the same audio with the first 10 seconds removed as the comparison buffer. Each implementation was run 50 times in shuffled order after decoding the audio once.

Branch	Mean execution time (s)
main	4.664295812019845
this branch	0.6442086398595711

The audio file is available on Wikimedia Commons:
https://commons.wikimedia.org/wiki/File:Pyotr_Ilyich_Tchaikovsky_-_1812_overture.ogg

[comavius]

@stephenjolly
Could you please review this PR? It speeds up hot-path algorithm in both time complexity order and benchmarking score. There's no breaking change in function interfaces and invariants.

[chrisn]

Thanks for this, looks like a nice improvement! The CI tests were failing, so I've just fixed those. Would you mind rebasing this branch on top of the latest master branch so we can verify all the checks are passing?