Guard the datashader import in the rasterizer benchmark

[brendancol]

Closes #3452.

benchmarks/rasterizer_benchmarks.py keeps datashader as a deliberate A/B comparison, so it stays referenced. But once datashader is no longer installed by default, the unconditional import datashader as ds made the whole script fail to run. This wraps that import (plus spatialpandas, which only feeds the datashader path) in a try/except.

• With datashader present, HAS_DATASHADER is True and the full comparison runs unchanged.
• With it absent, HAS_DATASHADER is False, the polygon path skips the datashader timing run, and the datashader cells show -- (same as how the cupy column already handles a missing backend).

Backend coverage: not applicable, this is a benchmark script rather than a spatial op.

Test plan:

• With datashader installed: confirmed HAS_DATASHADER is True and datashader takes part in both the timing and consistency runs.
• With datashader and spatialpandas blocked at import: confirmed HAS_DATASHADER is False, the script imports, the xrspatial/geocube/rasterio runners and consistency checks still work, and the markdown report builds.
• py_compile passes.
• No packaged unit test: the standalone benchmark scripts have no pytest coverage, and importing this one pulls in benchmark-only deps (geocube, spatialpandas) that the CI test env does not have.

https://claude.ai/code/session_01TF3CsJ7nee9agjajEWLw5d

[brendancol]

PR Review: Guard the datashader import in the rasterizer benchmark

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• benchmarks/rasterizer_benchmarks.py (generate_markdown): when datashader is absent, the polygon timing and ratio tables still print a "datashader" column header filled with --. The issue wording says "skip the datashader column". Dropping the column outright for polygons would mean threading HAS_DATASHADER into generate_markdown. The -- fallback matches how the cupy column already handles a missing backend, so I read this as a reasonable consistency choice rather than a defect. Flagging it so the maintainer can decide.
• benchmarks/rasterizer_benchmarks.py:39: the guard catches only ImportError. If datashader is installed but breaks on import for some other reason, that will propagate. Standard practice and fine; noting for completeness.

What looks good

• spatialpandas and datashader are folded into one guard. That is correct, since spatialpandas only feeds the datashader path, so there is no half-guarded state.
• Every reference to ds. and spatialpandas (the run_datashader body, the spd_gdf build) sits behind use_datashader, which now requires HAS_DATASHADER. No NameError path when datashader is missing.
• use_datashader = kind == "polygon" and HAS_DATASHADER is the minimal correct gate.

Checklist

• Algorithm matches reference/paper: n/a (benchmark infra)
• All implemented backends produce consistent results: n/a
• NaN handling is correct: n/a
• Edge cases covered: both datashader-present and datashader-absent paths verified
• Dask chunk boundaries handled correctly: n/a
• No premature materialization or unnecessary copies: yes
• Benchmark exists or not needed: this is the benchmark
• README feature matrix updated: not applicable, no new function
• Docstrings present and accurate: the guard carries an explanatory comment