add cent decomp

[cameroncuster]

lots of decisions going on in this one, so I'm looking for feedback :)

to explicitly list a few I'm looking for feedback on

• ripping the fft code
• using a trait for the cent_decomp
• documenting params (it felt useful in this particular case)
• the variable names (was going fast)

[cameroncuster]

looks like the naive convolution is too slow on the github action machines even though it only takes 10 seconds on my laptop :/

I'll have to come up with something else... might have to RIP some code from another lib for FFT for now... TBD (I'll think about various ideas)