Revisit current filtered FFT implementation

[daurer]

with recent updates to cupy and nvidia math library, it might be possible to avoid the pre-compiled filtered cufft.

[bjoernenders]

What we want to use here is this: the compile-prolog and compile-epilog. Esentially we only have to to use the FFT constructor for the propagator class and create functions fro post-fft and pre-fft to compile. It will add another dependency but it would save us the precompiled fft which was an awkward construct to begin with. This can actually be a nice exercise for somebody wanting to do something cuda