Add CI performance benchmarks with regression detection

[brendancol]

Problem

The benchmarks/ directory exists with ASV configuration and benchmark files for 13 modules, but:

1. Benchmarks are never run in CI — there is no performance regression detection.
2. Benchmark results may be stale.
3. No coverage for recently added features (dask viewshed, cost_distance dask, emerging_hotspots).
4. The CI matrix (.github/workflows/test.yml) runs pytest on 3 OS × 3 Python versions but has no performance tracking.

For a library whose primary value proposition is performance (numba JIT, GPU, dask), not detecting regressions is a critical gap.

Proposed Fix

1. Add an ASV-based benchmark CI job on a fixed Ubuntu runner (consistent hardware for comparison).
2. Create/update benchmarks for the top 5 performance-sensitive functions: slope, proximity, zonal.stats, cost_distance, focal_stats.
3. Each benchmark should test at least numpy and dask backends at representative sizes (e.g., 2000×2000 for numpy, 10000×10000 for dask).
4. Store benchmark results as CI artifacts and compare against the base branch.
5. Add a performance label to PRs that touch numba/CUDA kernels for automatic benchmark runs.
6. Fail CI on >20% regression (configurable threshold).

Impact

A library that bills itself as "fast" and "scalable" needs to prove it on every commit. Without regression detection, a well-intentioned refactoring could introduce a 10× slowdown in a numba kernel and nobody would notice until users complain.