Switch FFT backend to vkfft

silx.math.fft works reasonably well but has the following drawbacks:

• silx.math creates unnecessary contexts (fixed in recent version that will take months to be released)
• OpenCL: gpyfft is cumbersome to install, and its maintenance state is not clear
• Cuda: CUFFT memory usage is huge (3.5-4X the original data)
• Cuda scikit-cuda also sometimes creates unnecessary contexts

It might be good to switch to pyvkfft which addresses all these shortcomings. However, we have to ensure it works on all ESRF machines - for now there is an issue with nvrtc in several cases.