Add multiple CUDA streams support

[mxmlnkn]

Currently there seems to be no support for cufftSetStream.

[Flamefire]

This is correct. The reason for this is, that only a common subset of all libraries functionality can be used. There is no such thing as streams in the MKL or FFTW so I did not want to introduce something library specific like streams. If you have an idea how to cleanly add this, I'd be grateful.

[mxmlnkn]

Well, the includes are already library specific, so as a workaround may you could use the same preprocessor branching for setting the stream:

#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
#   include "foobar/libraries/cuFFT/cuFFT.hpp"
#   define makeFftPlan(...) makeFFT< libraries::cuFFT::CuFFT<> >( __VA_ARGS__ )
#else
#   include "foobar/libraries/fftw/FFTW.hpp"
#   define makeFftPlan(...) makeFFT< libraries::fftw::FFTW<> >( __VA_ARGS__ )
#endif
[...]
#ifdef ALPAKA_ACC_GPU_CUDA_ENABLED
    cufftSetStream( ftPlan, rStream );
#endif

The problem here is that ftPlan would need to be the raw plan, not the abstraction provided by HaLT.

[Flamefire]

What if I do expose a setStream method on the plan? It could be an empty default implementation but possibly overwritten for e.g. CUDA. For other backends it might be something different. Or a method to get the plan itself. Could be useful for fine-tuning of other things too, although this is against the single-interface design of the library where you just change the backend at one place (instead of the define you can just put a using FFT_Backend = libraries::fftw::FFTW<> there)

[mxmlnkn]

A setStream method sounds good to me.