Pure xtensor FFT implementation

[spectre-ns]

Checklist

• [x] The title and commit message(s) are descriptive.
• [x] Small commits made to fix your PR have been squashed to avoid history pollution.
• [x] Tests have been added for new features or bug fixes.
• [x] API of new functions and classes are documented.

Description

[spectre-ns]

@tdegeus Try again on the FFT PR. :)

[spectre-ns]

@JohanMabille I thought you might like this to help fill out the numpy interface farther.

[JohanMabille]

Thanks! We also have FFTW bindings, although not maintained for a while. It might be interesting to compare the results and performance.

[spectre-ns]

We also have FFTW bindings, although not maintained for a while. It might be interesting to compare the results and performance.

@JohanMabille Yeah I use the FFTW binding in some production code but it's nice to get something for prototyping without the extra dependencies. Also, FFTW is GPL licensed which meant we had to link against MKL instead.

I expect the performance from FFTW will be far superior because I didn't implement any memory pools or plans. Thus, there is dynamic memory allocation each time this is called in the hot path. We could implement it such that when used with xfixed the operation is completely fused and unwrapped which would have an advantage over FFTW... FFTW is crazy fast though so I don't know if xtensor would win versus the hand rolled implementation in FFTW.