Improve performance of shifted ApplyMasksUDF

[uellue]

At the moment the shift compensation in ApplyMasksUDF can be a lot slower than the uncompensated one. When testing with the Radial Fourier example notebook and sparse masks, it was a bout a factor 25 slower. Dense masks will lead to an out of memory error there.

With optimized kernels that apply shifts to index calculations instead of slicing data, better performance should be achievable. Perhaps only a factor 3 slower?