Review `nrefocus.refocus_stack` and `nrefocus.autofocus_stack` to see if they are redundant due to #25

[PinkShnack]

Now that nrefocus accepts 3D arrays, it could be that nrefocus.refocus_stack and nrefocus.autofocus_stack are redundant. They use multiprocessing within their functions to be able to process a stack of 3D data.

[paulmueller]

Does numpy natively circumvent the GIL when computing the Fourier transform on 3D arrays? It would be interesting to see whether native numpy is faster than the multiprocessing approach.

nrefocus.refocus_stack uses multiprocessing.Pool.map_async which looks very efficient. But if FFTW is faster than that, then the numpy implementation can be considered "for educational purpose only", and I would be glad to deprecate nrefocus.refocus_stack.