BUG: ndimage: fix origin handling for {minimum, maximum}_filter

[grlee77]

Reference issue

closes #20652

What does this implement/fix?

fixes incorrect assumption that len(origin) should match ndim. It should have instead been len(axes) with a value of 0 inserted afterwards for any non-filtered axis. This same solution was already previously implemented for rank_filter.

New test cases were added that fail with the error mentioned in #20652 prior to this change (they did not fail previously for rank_filter, median_filter or percentile_filter, but it was easy to add those to the test parameters).

[mdhaber]

I understand most of this since I reviewed those PRs. I see that the fix is the same as what was already done for rank_filter. The test, which makes sense, does fail in main for the minimum/maximum filters but not the others, as expected. So nothing is jumping out at me - but my eyes aren't very sensitive in this module. Would you like for this to be reviewed by someone else who spends more time with this code (or at lest another set of eyes, since I didn't notice the bug originally)?

[grlee77]

Would you like for this to be reviewed by someone else who spends more time with this code (or at lest another set of eyes, since I didn't notice the bug originally)?

Thanks for checking it. I think the test case covers the issue well enough so that additional review is probably not needed.

Apparently this is a pretty uncommon code path as no bug has been raised for it since the 1.11 release

[tylerjereddy]

I poked around locally and found the same thing as Matt. Sounds like we're in favor of this and fails before/passes after, so I'll put it in.

I'm sometimes a bit nervous about the exact test comparisons on floating-point type arrays like this. I'll assume that those really are guaranteed to be identical, but if they aren't that's a small follow-up fix.

[tylerjereddy]

thanks both