Forward per-tool repr to the accessed tool (#3478)

[brendancol]

Summary

Closes #3478.

After #3477 added a categorized __repr__ to the .xrs accessor, evaluating a single tool without calling it (e.g. da.xrs.slope) printed the whole accessor catalog. da.xrs.slope is a bound method, and Python builds a bound method's repr as <bound method Cls.name of {repr(self)}>, so repr(self) routed through the accessor's catalog __repr__ and embedded the entire listing.

• Accessor attribute access now wraps public bound methods in a small callable _AccessorTool proxy. Its __repr__ / _repr_html_ show that tool's own signature and docstring instead. Calls forward unchanged, and __doc__ / __wrapped__ are mirrored so help() and inspect keep seeing the delegated docstring.
• The accessor-level catalog repr (da.xrs) is untouched.

Backend coverage

Pure accessor-layer change; no compute path is touched. Behaves the same across numpy / cupy / dask+numpy / dask+cupy.

Test plan

• repr(da.xrs.slope) shows slope's signature + docstring, not the catalog
• tools stay callable and match the direct function call
• inspect.getdoc / help() metadata preserved on the proxy
• HTML repr scoped to the tool
• Dataset accessor scoped the same way
• accessor-level catalog repr still renders
• full test_accessor.py suite passes (86 tests), flake8 clean

[brendancol]

PR Review: Forward per-tool repr to the accessed tool (#3478)

I read accessor.py in full and exercised the proxy against the edge cases below. The fix is correct and targeted. No blockers.

Blockers

None.

Suggestions

None.

Nits (optional)

• _AccessorTool returns a fresh proxy on every attribute access, so da.xrs.slope == da.xrs.slope is now False where a bound method compared equal. Harmless for normal use, but anything downstream comparing accessor methods by equality would see the change. Low risk; not worth caching unless it bites.
• __getattribute__ now runs for every attribute access on the accessor, including self._obj inside each method body. The early name.startswith("_") return keeps the hot path cheap (one extra call plus a startswith), so this is fine, just noting the cost is non-zero.

What looks good

• Root cause is correctly identified: a bound method's repr embeds repr(self), which routed through the catalog __repr__. Wrapping at attribute access is the right layer.
• The proxy mirrors __doc__ / __wrapped__ / __name__, so help(), inspect.signature, and inspect.getdoc keep working. Verified inspect.signature(da.xrs.slope) returns (**kwargs) and getdoc matches the standalone function.
• Private attributes and the catalog __repr__ are left untouched, so da.xrs still renders the full listing.
• Calls forward unchanged, positional args included (ndvi, rasterize are covered by the existing suite, which now all runs through the proxy).
• Missing attributes still raise AttributeError.
• Tests cover the scoped text repr, HTML repr, callable forwarding, help metadata, the Dataset accessor, and the unchanged catalog repr. Both accessors are handled.

Checklist

• Root cause correctly identified
• help()/inspect metadata preserved
• Catalog repr unchanged
• Calls forward unchanged (incl. positional args)
• Both DataArray and Dataset accessors covered
• Tests present and passing (86 in test_accessor.py), flake8 clean
• No compute path touched; backend-agnostic