Add categorized repr to the .xrs accessor (#3476)

[brendancol]

Closes #3476

Printing the .xrs accessor used to show the default object repr, which
tells you nothing about what you can call. Now it lists the available
operations grouped by category, so you can see the toolset from a plain
REPL or notebook without tab completion or an LSP.

• __repr__ renders a grouped, wrapped text listing; _repr_html_
  renders a compact table in Jupyter.
• The grouping is parsed from the # ---- Category ---- banners already
  in the accessor source, so it can't drift from the actual method set. A
  test asserts every public method lands in exactly one category.
• Falls back to a flat alphabetical method list if source introspection
  isn't available (e.g. a frozen build).

Example:

<XrsSpatialDataArrayAccessor> xarray-spatial tools for this DataArray
call as: .xrs.<name>(...)

Surface:
    slope, aspect, hillshade, curvature
Hydrology:
    flow_direction, flow_accumulation, watershed, ...
Multispectral (self = NIR band):
    ndvi, evi, arvi, savi, nbr, sipi
...

Backend coverage: not applicable. This is a pure-Python repr on the
accessor classes; it touches no array code, so numpy / cupy / dask paths
are unaffected.

Test plan:

• pytest xrspatial/tests/test_accessor.py (79 passed)
• New tests: category coverage of all public methods, text and HTML
  repr content, DataArray vs Dataset distinct listings, HTML escaping,
  source-unavailable fallback
• flake8 clean on changed files

[brendancol]

PR Review: Add categorized repr to the .xrs accessor (#3476)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• The text and HTML fallbacks label the no-category case differently. The
  text path prints "(categories unavailable; flat method list)"; the HTML path
  uses an "Available tools" row header (accessor.py:638 vs accessor.py:655).
  Harmless, but matching the wording would make the two paths read the same.

Nits (optional improvements)

• accessor.py:619: _accessor_categories is lru_cached and keyed on the class.
  Only two classes ever reach it, so the cache is a two-entry memo. Fine as-is;
  noting that maxsize=None is unbounded by design, not an oversight.

What looks good

• The parser reads the # ---- Category ---- banners already in the source, so
  the listing can't drift from the real method set. The coverage test
  (test_categories_cover_every_public_method) asserts an exact match against
  vars(cls), which also catches a future nested non-underscore def leaking into
  the listing.
• HTML output escapes method and category names, with a dedicated test that
  injects a <script> name to prove it.
• The source-unavailable fallback is tested with a type()-built class whose
  source inspect can't retrieve.
• Pure-Python change with no array code, so no backend / dask / cupy / NaN
  concerns apply.

Checklist

• Algorithm matches reference/paper: N/A (no algorithm)
• All implemented backends produce consistent results: N/A (no array code)
• NaN handling is correct: N/A
• Edge cases are covered by tests: yes (fallback, HTML escaping, DataArray vs Dataset)
• Dask chunk boundaries handled correctly: N/A
• No premature materialization or unnecessary copies: N/A
• Benchmark exists or is not needed: not needed (repr is interactive, not a hot path)
• README feature matrix updated: N/A (no new function)
• Docstrings present and accurate: yes (helpers documented; quickstart updated)

[brendancol]

Follow-up review (after e6d1be2)

The one Suggestion from the first pass is resolved. The text and HTML reprs
now both route through _categories_or_fallback (accessor.py:644), so the
no-banner case shows the same "Available tools" group in either path. The
fallback test was updated to assert the shared label.

No new findings. The Nit about lru_cache(maxsize=None) was a note, not a
defect, and stands as-is (two classes, intentional). Pure-Python repr change;
no backend, dask, cupy, or NaN concerns apply. Nothing blocking.