Make PolyDiff work with variable step size and missing data

[pavelkomarov]

In principle a polynomial fit can be performed to data spaced wherever, and NaNs can simply be ignored. But np.polyfit loses its mind and returns all NaNs if given a NaN input, and the way slide_function is implemented presently doesn't smoothly account for data points that might be nonuniformly spaced. These are both completely addressable, but in practice require some thinking, NaN-checking, and probably some light refactoring. Cursor can probably come up with some useful ideas, but it may still take a human hand to properly integrate and polish, especially because polydiff isn't the only one dependent on slide_function or the kernel functions.

This one may be more complicated still, because as I write in the taxonomy paper's last table, "When frequency of points can vary, optimal degree and window size can fluctuate across domain." That could make this thing a mess to try to parameterize and optimize in the general case, or make it so that we're significantly under-parameterizing and under-exploring if we enforce common choices across the domain. For slight nonequispacedness this wouldn't matter too much, though, and it would be nice if it just worked$^\text{TM}$.

[pavelkomarov]

Some ideas from Cursor. Not sure I totally agree with its design choices, but this change will reach a couple layers of utility functions deep.