scipy.fft.fft2 is 2-3x faster than np.fft.fft2

One of the bottlenecks of the local coregistration is the fourier transform. If pyfftw is not available (which is everywhere, because the most recent version is currently disabled), numpy is used.

In my testing (M3 Mac), the scipy.fft.fft2 implementation is 2-3x faster than np.fft.fft2 within arosics.

Time per function call
479822.5  fft_arr0 = np.fft.fft2(in_arr0)
350634.3  fft_arr1 = np.fft.fft2(in_arr1)

171587.7  fft_arr0 = scipy.fft.fft2(in_arr0)
154358.3  fft_arr1 = scipy.fft.fft2(in_arr1)

It's a drop-in replacement in CoReg.py, scipy is already used elsewhere.