convolution vs correlation: kernel flip convention
Maybe it's me, but I feel our convolutions still aren't proper convolutions:
The kernel is not flipped before it is multiplied by the image: so the result is, in turn, flipped.
This does not affect the PSF reconstruction, but it lays the burden on the user to flip the narrow PSF obtained with starred before proceeding I think. Maybe we can flip it in our two main convolution functions? (`pad_and_convolve` and pad_and_convolve_fft
)