Propagate input dims/coords through true_color() (#3429)

[brendancol]

Closes #3429.

true_color() hardcoded its output dims to ['y','x','band'] and rebuilt coords via r['y']/r['x']. That raised KeyError: 'y' on rasters with lat/lon (or row/col) dims and silently dropped non-spatial coords like spatial_ref. The other indices in the module already pass dims/coords/attrs straight through; this brings true_color() in line.

• Derive the two spatial dims and their coords from the input DataArray, then append the band dim.
• Pass input attrs and any extra coords through unchanged.

Backend coverage: numpy, cupy, dask+numpy, dask+cupy (all verified locally on a GPU host).

Test plan:

• New test_true_color_preserves_non_yx_dims runs on all four backends with lat/lon input.
• New test_true_color_preserves_extra_coords checks a spatial_ref coord survives.
• Full test_multispectral.py passes (159 tests).

[brendancol]

PR Review: Propagate input dims/coords through true_color() (#3429)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• multispectral.py:1828 — if an input raster carried a coord literally named band, the 'band': [0, 1, 2, 3] entry would overwrite it. Far-fetched (a 2D raster with a non-dim coord called band), and the old code had the same collision, so it's not a regression. Not worth guarding against unless it ever comes up.

What looks good

• The fix mirrors how every other index in the module preserves metadata, so true_color is no longer the odd one out.
• Metadata is built in the public function after dispatch, so all four backends share one code path. The parametrized test runs numpy, cupy, dask+numpy, and dask+cupy with lat/lon input.
• [*r.dims, 'band'] lines up with the (H, W, 4) output cube: r is validated 2D and the stack axis is appended last on every backend.
• Tests assert on dims, coord values, and attrs, not just that the call returns. The spatial_ref test covers the dropped-coord half of the bug.

Checklist

• Algorithm matches reference/paper (N/A, metadata-only change, pixel math untouched)
• All implemented backends produce consistent results
• NaN handling is correct (unchanged)
• Edge cases are covered by tests (lat/lon dims, extra coord)
• Dask chunk boundaries handled correctly (metadata built post-dispatch, no chunk interaction)
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (not needed, no perf change)
• README feature matrix updated (N/A, no new function or backend change)
• Docstrings present and accurate