Rasterize LinearRing as a line; warn on dropped geometries

[brendancol]

Closes #3055

_classify_geometries sorts inputs into polygon / line / point buckets. The line bucket matched only LineString (1) and MultiLineString (5), so a non-empty LinearRing (shapely type id 2) fell through every bucket and disappeared. rasterize then returned an all-fill raster with no error or warning.

• Add type id 2 to the fast-path line bucket and the 'LinearRing' name to the slow-path (GeometryCollection) line branch, so a LinearRing rasterizes as its closed boundary.
• Warn, naming the type, whenever a non-empty geometry matches no bucket, so data is never dropped silently. None / empty geometries stay silent (None reports as non-empty with type id -1, so it is now treated as empty to match the slow path's existing guard).
• Update the rasterize docstring to list LinearRing and the new drop-with-warning behavior.

_classify_geometries is the single classifier shared by every backend path, so the fix applies to numpy, cupy, dask+numpy, and dask+cupy.

Test plan

• LinearRing routes to the line bucket on both the fast path and the GeometryCollection slow path
• End-to-end rasterize burns a LinearRing boundary and matches the equivalent closed LineString (numpy, dask, cupy, dask+cupy)
• Non-empty unsupported type ids warn; None and empty geometries stay silent
• Full existing rasterize suite passes (586 passed, 2 skipped); flake8 clean

[brendancol]

PR Review: Rasterize LinearRing as a line; warn on dropped geometries

Blockers (must fix before merge)

• None.

Suggestions (should fix, not blocking)

• None blocking. The else branch in _classify_one (slow path, rasterize.py:366) is effectively unreachable through the public API, since shapely won't build a GeometryCollection that holds an unsupported member. It is reasonable as defense-in-depth, and the helper itself is unit-tested, so I would keep it.

Nits (optional improvements)

• _warn_dropped_geometries warns with the default UserWarning. That matches the rest of the module (no explicit warning category is used anywhere in rasterize.py), so no change is needed. Noting it only so the choice is visible.

What looks good

• LinearRing now routes to the line bucket on both the vectorized fast path (rasterize.py:308) and the GeometryCollection slow path (rasterize.py:363), and the tests confirm a ring burns the same pixels as the equivalent closed LineString.
• The None case is handled correctly: valid = ~empty & (type_ids >= 0) (rasterize.py:300) drops type id -1, so a None geometry is skipped silently, matching the slow path's existing geom is None guard instead of warning on it.
• The drop warning fires once per batch with the unique offending types, not once per geometry, so it will not spam.
• Backend coverage is real: _classify_geometries is the single classifier shared by numpy, cupy, dask+numpy, and dask+cupy, and the tests exercise all four end to end.

Checklist

• Algorithm matches intended behavior (LinearRing = closed line)
• All implemented backends produce consistent results
• NaN / None / empty handling is correct
• Edge cases covered (None, empty, unsupported type id, GC slow path)
• Dask chunk boundaries handled (shared classifier; ragged chunks tested)
• No premature materialization or unnecessary copies
• Benchmark not needed (classification-only bug fix)
• README feature matrix unchanged (no new function, no backend change)
• Docstrings present and updated