Fix crop for WCSes with interleaved world objects

[nabobalis]

PR Description

crop takes one entry per world object, with None for objects that aren't cropped. It ordered the objects by where each one last appears in the WCS, but pixel_to_world and world_to_pixel use where each first appears. So on WCSes with interleaved world objects (e.g. HPLN, WAVE, HPLT, or DKIST VISP), crop rejected points in pixel_to_world order, including partial ones. The error for a point of the wrong length also didn't say what it expected.

import numpy as np

from astropy.wcs import WCS

from ndcube import NDCube

w = WCS(naxis=3)
w.wcs.ctype = ["HPLN-TAN", "WAVE", "HPLT-TAN"]
w.wcs.cunit = ["deg", "m", "deg"]
w.wcs.cdelt = [0.1, 1e-10, 0.1]
w.wcs.crpix = [1, 1, 1]
w.wcs.crval = [0, 5e-7, 0]
w.wcs.dateobs = "2020-01-01"
w.wcs.set()
cube = NDCube(np.zeros((4, 5, 6)), wcs=w)
sky0, wl0 = cube.wcs.pixel_to_world(1, 1, 1)
sky1, wl1 = cube.wcs.pixel_to_world(3, 3, 2)

for label, points in {
    "[sky, wl]": ([sky0, wl0], [sky1, wl1]),
    "[sky, None]": ([sky0, None], [sky1, None]),
    "[None, wl]": ([None, wl0], [None, wl1]),
    "[wl, None]": ([wl0, None], [wl1, None]),
    "[wl, sky]": ([wl0, sky0], [wl1, sky1]),
    "[sky]": ([sky0], [sky1]),
}.items():
    try:
        result = cube.crop(*points).shape
    except Exception as e:
        result = f"{type(e).__name__}: {e}"
    print(f"{label:12} -> {result}")

points	main	this PR
[sky, wl] (pixel_to_world order)	TypeError: <class '...SkyCoord'> of component 0 in point 0 is incompatible with WCS component spectral <class '...Quantity'>.	(2, 3, 3)
[sky, None]	same TypeError	(2, 5, 3)
[None, wl]	TypeError: <class '...SpectralCoord'> of component 1 in point 0 is incompatible with WCS component celestial <class '...SkyCoord'>.	(4, 3, 6)
[wl, None]	ValueError: Expected the following order of world arguments: SkyCoord	(4, 3, 6)
[wl, sky] (old order)	(2, 3, 3)	(2, 3, 3)
[sky]	ValueError: 1 components in point 0 do not match WCS with 2 components.	same, followed by Each point must have one entry per world object (use None for a component that should not be cropped), in order: celestial (SkyCoord), spectral (Quantity).

Changes:

• Objects are ordered by first appearance (utils.misc.unique_sorted, as in axis_world_coords).
• If each non-None object matches exactly one world object class, objects are matched by class, as in astropy's world_to_pixel. Otherwise they are matched by position.
• The length and type errors list the expected objects in order (replaces Name expected world objects in crop component-count error #939).
• get_crop_item_from_points sorts the pixel axes instead of using set order, which swapped some bounds on cubes with nine or more dimensions.

User-visible changes:

• Points in pixel_to_world order work, and so does the [spec, sky] example in Allow greater flexibility in crop bounds order #608.
• Points in the old order still work when every world object has its own class (none a subclass of another), as on DKIST VISP. dkist's crop tests pass, two of them only because of the class matching.
• Matching is all or nothing for each point. If classes are shared (FITS WAVE and LINEAR are both Quantity) or one is a subclass of another (SpectralCoord and Quantity), the point is matched by position, in first-appearance order. So the old-order point [wl, sky, None] on HPLN, WAVE, HPLT, LINEAR now raises.
• Some inputs that main rejected with a TypeError now land in the wrong slot:
  • an object of the same class as another slot, given in that slot's place (as with world_to_pixel)
  • a plain Quantity for a gWCS SpectralCoord slot, which moves to the only Quantity slot
  • with wcs=cube.extra_coords, a Quantity of the wrong type, which moves to a dummy pixel slot (crop(wcs=cube.extra_coords): dummy pixel components leak into points and errors #978)

AI Assistance Disclosure

AI tools were used for:

• Code generation (e.g., when writing an implementation or fixing a bug)
• Test/benchmark generation
• Documentation (including examples)
• Research and understanding
• No AI tools were used

Regardless of AI use, the human contributor remains fully responsible for correctness, design choices, licensing compatibility, and long-term maintainability.

K