Optimizations inspired by ssqueezepy

[cboulay]

ssqueezepy has a bunch of optimizations around FFT and CWT.

• numba jit
• fftw if available
• optional parallel processing for fft
• torch

jit

I tried this once and it was slower than what I had already. However, it's possible that I was using something that number couldn't handle and it fell back to python. I should try again with nopython=True to make sure I'm not providing numba-incompatible code.

fftw

This should be an easy win. The first sample will take much longer to process while fftw does its optimization but subsequent processing will be faster.

parallel

scipy.fft(..., workers=N) -- does this actually help? Isn't the overhead more than the savings?

torch

We will need all the nodes in a standard pipeline to handle torch tensors before the overhead of moving to gpu/mps is worth it.