Fix performance regression in Coordinates.to_index

[thodson-usgs]

Summary

Fixes #11305.

The codes passed to pd.MultiIndex in Coordinates.to_index were being converted from the cache-friendly ndarrays produced by np.tile/np.repeat into Python lists. That conversion was introduced in #10694 to silence a mypy arg-type error (pandas-stubs declares codes: Sequence[Sequence[int]]). The per-element materialisation dominates runtime for large indexes.

This PR passes code_list directly and suppresses the mypy complaint the same way we already do for levels on the line above (# type: ignore[arg-type,unused-ignore]).

Verified that pd.MultiIndex with ndarray codes produces a result that is .identical() to the list-of-int version (same internal code dtype, same values), so the change is purely a performance fix.

MVCE (from #11305)

# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "xarray[complete]@git+https://github.com/pydata/xarray.git@refs/pull/11306/head",
# ]
# ///
import time
import numpy as np
import xarray as xr

da = xr.DataArray(np.arange(100 * 2000 * 300).reshape(100, 2000, 300), name="foo")
t = time.perf_counter()
da.to_dataframe()
print(f"to_dataframe: {time.perf_counter() - t:.3f}s")

Run with uv run script.py. Local results on my machine:

	to_dataframe
main (before)	10.33s
this PR (after)	0.26s

~40× speed-up, matching the ~10s reporter saw.

Test plan

Ran locally with uv run pytest -n auto:

• xarray/tests/test_coordinates.py — 27 passed
• xarray/tests/test_variable.py, test_concat.py, test_missing.py — 827 passed, 69 skipped, 9 xfailed, 3 xpassed
• xarray/tests/test_dataset.py, test_dataarray.py, test_groupby.py — 1458 passed, 91 skipped, 5 xfailed, 2 xpassed
• pre-commit clean on touched files, no new mypy errors introduced on xarray/core/coordinates.py

I grepped for callers of Coordinates.to_index (xarray/core/{dataset,dataarray,variable,indexes,missing,accessor_dt}.py, xarray/plot/facetgrid.py, xarray/groupers.py) — all paths produce a pd.MultiIndex whose consumers rely on .values / .get_loc / iteration, which are unaffected by whether codes was built from ndarrays or lists.

---

[This is Claude Code on behalf of Tim Hodson]

[welcome]

Thank you for opening this pull request! It may take us a few days to respond here, so thank you for being patient.
If you have questions, some answers may be found in our contributing guidelines.

[thodson-usgs]

cc @max-sixty (introduced the codes=[list(c) for c in code_list] conversion in #10694 to silence a mypy arg-type error) and @ianhi (reverted the analogous levels conversion in #10705 for dtype reasons, left the codes one in place) — curious if either of you has context on why the list(...) conversion was needed for codes specifically beyond the mypy stub, or any concern with the suppress-via-type-ignore approach used here.

[This is Claude Code on behalf of Tim Hodson]

[ianhi]

and @ianhi (reverted the analogous levels conversion in #10705 for dtype reasons, left the codes one in place) —

I strongly suspect this was an oversight on my part - or rather that I wasn't seeing an immediate error from it so I didn't fix it

[thodson-usgs]

@ianhi , I ran the script to verify the performance improvement. Since this is a triage one-liner, I'm going to open the PR. Thanks for your feedback.