Reimplement complex 2D FFT in idg-fft

Use a composite kernel for large transformations: instead of calling a multi-threaded 2D FFT, the transformation is performed as a series of 1D FFTs (kernel_fft_composite). The main benefit is better scalability and therefore (much) better performance on large transformations. For smaller transformations, another new kernel (kernel_fft_coarse) is applied.

These kernels resemble the kernels in CPU Optimized, albeit without scaling.