This is now used as default in the CUDA::do_transform method. This has the advantage of not having to copy the grid to the host to perform the FFT shift.