Support N-D table coords, subclass attr propagation, clearer crop error

[nabobalis]

PR Description

Allow a single N-D Time or Quantity lookup table in extra_coords, representing one coordinate varying over several pixel dimensions. Add NDCube._extra_attrs_to_copy so subclasses can declare attributes propagated through arithmetic and to_nddata. Make the crop component-count error name the expected world objects.

Description

This PR upstreams three improvements motivated by irispy-lmsal's gWCS raster
work, where IRIS spectrograph observations are represented as 4-D
(scan, step, slit, wavelength) cubes. Each change removes a workaround that
instrument packages currently have to carry themselves.

1. N-D Time/Quantity lookup tables in extra_coords

TimeTableCoordinate and QuantityTableCoordinate now accept a single N-D
table, representing one physical coordinate that varies over N pixel
dimensions. For example, a 2-D Time table indexed by (raster scan, raster
step) can now be attached with:

cube.extra_coords.add("time", (0, 1), times_2d, physical_types="time")

Previously QuantityTableCoordinate raised "Currently all tables must be
1-D" and TimeTableCoordinate was hardwired to one dimension, so per-exposure
times on a multi-scan cube could not be expressed as an extra coord at all.

Implementation notes:

• N-D tables expose their model inputs in pixel order (reversed array
  order). NDCube._generate_world_coords transposes world values assuming
  the APE-14 pixel-axis convention, so an array-ordered model produces
  transposed axis_world_coords output. ExtraCoords.mapping reverses the
  converted axes for such coordinates to stay consistent (see the new
  BaseTableCoordinate._model_inputs_are_pixel_ordered property).
• ExtraCoords._getitem_lookup_tables now drops integer-sliced axes from a
  table's axes tuple when the sliced table actually loses pixel dimensions.
  Previously the axes tuple kept stale (even negative) entries; this was
  latent for existing coordinate types but breaks N-D tables. Range slices
  keep the table N-D, integer slices reduce it, and slicing away all table
  dimensions drops the coordinate into dropped_tables as before.
• interpolate supports N-D tables via scipy.interpolate.interpn
  (TimeTableCoordinate.interpolate now takes *new_array_grids, matching
  the other table coordinate types; the old single-grid call still works).
• Fixed a latent Length1Tabular dispatch bug: a table of shape (1, n) was
  routed to the length-1 1-D model because the check used len(table) == 1
  instead of table.shape == (1,).

2. NDCube._extra_attrs_to_copy extension point

Subclasses that carry domain-specific instance attributes (observer
coordinates, per-step WCS tables, instrument flags, ...) currently have to
override _new_instance and to_nddata to keep those attributes from
silently disappearing on every derived cube. Forgetting one call site means
arithmetic or a utility function returns a cube with the metadata gone and no
error.

Subclasses can now declare:

class MyCube(NDCube):
    _extra_attrs_to_copy = ("observer", "calibration_level")

and the attributes are propagated (by reference) through _new_instance
(arithmetic, __neg__, fill_masked, ...) and through to_nddata when the
target nddata_type is a subclass carrying the same attributes. Explicit
kwargs still override the automatic copy, and plain NDData targets are
unaffected. Slicing is deliberately not covered: shape-dependent attributes
cannot be sliced generically, so subclasses keep handling that themselves
(documented on the attribute).

3. Crop component-count error names the expected world objects

Old code written against a previous WCS fails with
"2 components in point 0 do not match WCS with 4 components.", which gives
the user no clue what the four components are. The error now reads:

2 components in point 0 do not match WCS with 4 components. Each point must
have one entry per world object (use None for a component that should not be
cropped), in order: em (SpectralCoord), celestial (SkyCoord), time (Time),
custom_step (Quantity).

The existing message prefix is preserved, so downstream tests matching on it
keep passing.

Testing

• 13 new tests: N-D table construction, WCS round-trips, slicing (range,
  integer, full drop), interpolation, multi-N-D rejection, cube-level
  axis_world_coords through extra_coords including after slicing,
  attribute propagation through arithmetic and to_nddata, and the new crop
  error wording.
• Full ndcube suite passes locally (534 passed; one pre-existing
  environment-dependent failure with a newer gwcs,
  test_crop_by_extra_coords_values_all_axes_with_coord, fails identically
  on main).
• Validated against irispy-lmsal's gWCS raster branch: its full suite passes
  against this patch, and the 2-D (scan, step) exposure-time table that
  motivated change 1 attaches and slices correctly on a real 13-scan
  observation.

Notes for reviewers

• The pixel-ordering convention for N-D table models (point 1 above) is the
  part most worth scrutinizing: multi-1-D-table coordinates (e.g. meshed
  SkyCoord tables) keep their existing array-ordered convention, and only
  single N-D tables use the reversed order. An alternative would be to fix
  _generate_world_coords to consult the mapping instead, but that touches
  every WCS path rather than only the new feature.

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.

[nabobalis]

Honestly, this was created to make my life easier in creating a 5D gWCS. I don't even know if these changes make logical sense in the first place.