ENH: ndimage.vectorized_filter: generic_filter with memory, speed, and feature improvements

Hi Team,

gh-22575 proposes the addition of ndimage.vectorized_filter, a function similar to ndimage.generic_filter that takes advantage of vectorized Python callables. It is much faster than ndimage.generic_filter for moderate filter footprints and reasonably fast callables (up to 200x), and in fact, it can even be faster than the specialized filters like median_filter (for n-d input). It also allows the user to control memory usage, and it satisfies several other longstanding feature requests (e.g. `mode=‘valid’, multivariate filter output).

Thanks for considering the addition!
Matt