Fix NDCube.rebin for lookup-table extra coords

[nabobalis]

PR Description

Split out of #967. None of this depends on N-D lookup tables; it all reproduces on main and 2.4 with 1-D tables.

NDCube.rebin had three problems with lookup-table extra coords:

• They were resampled with offset=0, so each new pixel got the value at the first pixel of its bin, while the rebinned WCS gives the bin centre. rebin now passes (bin_shape - 1) / 2 for lookup tables. WCS-backed extra coords keep 0, because ResampledLowLevelWCS treats the offset as a pixel-edge shift. Nothing public tells the two apart, so this checks the private _lookup_tables.
• ExtraCoords.resample put the new grids in an object array. When they all have the same length (every 1-D cube, for example) numpy makes a 2-D object array and np.interp raises TypeError. They are now a list.
• TimeTableCoordinate.interpolate interpolated absolute MJD floats (about 0.6 µs resolution for current dates) and dropped reference_time. It now interpolates seconds from the first time and keeps reference_time.

test_rebin already expected bin-centre times, but np.allclose's default rtol=1e-5 is about half a day at MJD 51544, so the 30 s error passed. It now uses rtol=0, atol=1e-8.

The second commit drops the "New array grids must all be same shape" check from QuantityTableCoordinate.interpolate, so a mesh Quantity table works when its axes are rebinned to different lengths. It is separate in case you'd rather not take it.

For direct callers of TimeTableCoordinate.interpolate, or ExtraCoords.resample with kwargs, a numeric left/right is now seconds from table[0] rather than an absolute MJD, and left=right=np.nan gives a NaN Time instead of raising ValueError. rebin passes no kwargs.

import numpy as np
import astropy.units as u
from astropy.time import Time
from astropy.wcs import WCS

from ndcube import NDCube

wcs = WCS(naxis=2)
wcs.wcs.ctype = ["WAVE", "FREQ"]
wcs.wcs.cunit = ["m", "Hz"]

# Time table on array axis 1 at 60 s cadence, rebinned by 2 along that axis.
cube = NDCube(np.zeros((4, 6)), wcs=wcs)
cube.extra_coords.add("time", 1, Time("2026-01-01T00:00:00") + np.arange(6) * 60 * u.s)
r = cube.rebin((1, 2))
print(r.axis_world_coords(wcs=r.extra_coords)[0].isot)

# 1-D cube.
wave = WCS(naxis=1)
wave.wcs.ctype = ["WAVE"]
wave.wcs.cunit = ["m"]
cube1 = NDCube(np.zeros(6), wcs=wave)
cube1.extra_coords.add("wave2", 0, np.arange(6) * u.m)
r1 = cube1.rebin((2,))
print(r1.axis_world_coords(wcs=r1.extra_coords)[0])

main:

['2026-01-01T00:00:00.000' '2026-01-01T00:02:00.000'
 '2026-01-01T00:04:00.000']
Traceback (most recent call last):
  ...
TypeError: Cannot cast array data from dtype('O') to dtype('float64') according to the rule 'safe'

this branch:

['2026-01-01T00:00:30.000' '2026-01-01T00:02:30.000'
 '2026-01-01T00:04:30.000']
[0.5 2.5 4.5] m

Not in this PR:

• WCS-backed extra coords are still broken after rebin, for unrelated reasons: WCS-backed extra coords are unusable after NDCube.rebin #981.
• A non-mesh 2-D SkyCoord lookup table still collapses to its diagonal when rebinned (also on 2.4). I'll open a separate PR for that.

This adds one small conflict with #950 in ExtraCoords.resample; resolving it means keeping both changes.

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.