Add terrain metrics module: TRI, TPI, and Roughness

[brendancol]

Summary

• Adds three widely-used terrain morphometry metrics as 3×3 focal window operations on elevation values (no cellsize parameter needed):
  • TRI (Terrain Ruggedness Index) — sqrt(sum((z_neighbor - z_center)²)) over 8 neighbors (Riley et al. 1999)
  • TPI (Topographic Position Index) — z_center - mean(z_neighbors) (Weiss 2001)
  • Roughness — max(window) - min(window) in 3×3 window (GDAL definition)
• All four backends supported: NumPy, CuPy, Dask+NumPy, Dask+CuPy
• Boundary modes: nan, nearest, reflect, wrap
• @supports_dataset decorator for Dataset input
• .xrs accessor methods on both DataArray and Dataset
• ASV benchmarks following the existing Benchmarking pattern

Files changed

File	Change
xrspatial/terrain_metrics.py	New — core module with kernels, wrappers, public API
xrspatial/__init__.py	Add tri, tpi, roughness imports
xrspatial/accessor.py	Add methods to both accessor classes
xrspatial/tests/test_terrain_metrics.py	New — 179 tests
benchmarks/benchmarks/terrain_metrics.py	New — ASV benchmarks
README.md	Add 3 rows to Surface section

Test plan

• pytest xrspatial/tests/test_terrain_metrics.py -v — 179 passed
• Existing test_curvature.py still passes (75 passed)
• GPU backends verified on CUDA hardware — 48 passed (24 CuPy, 24 Dask+CuPy)
• Benchmarks run (ASV runner has a Python 3.14 compat issue, so ran directly)

Benchmark results

Metric	Size	NumPy (ms)	Dask (ms)	CuPy (ms)
TRI	100×50	0.10	36.27	0.39
TPI	100×50	0.11	10.10	0.40
Roughness	100×50	0.13	9.87	0.41
TRI	300×150	0.16	10.12	0.37
TPI	300×150	0.11	11.12	0.38
Roughness	300×150	0.14	11.14	0.34
TRI	1000×500	1.87	13.20	0.76
TPI	1000×500	0.60	12.76	0.35
Roughness	1000×500	1.22	13.35	0.33

Dask overhead dominates at small sizes; at 1000×500 the NumPy Numba JIT kernels are sub-2ms and CuPy stays under 1ms.