Implement SMRF ground detection (37% lower RMSE than our PMF)

[amarcozzi]

Summary

Add SMRF (Simple Morphological Filter, Pingel et al. 2013) as a ground-detection
algorithm alongside the existing PMF, implemented in-repo the way _pmf implements
Zhang et al. 2003, and validated against PDAL's filters.smrf.

This is an accuracy change, not a refactor. Measured on 9M points of vendor-classified
3DEP at 1 m, scoring each derived ground surface against the vendor's own ASPRS class 2 —
the nearest thing to truth this data has:

algorithm	time	RMSE	p95	coverage
ours (_pmf + snap)	0.54 s	0.056	0.020	92.7%
pdal filters.pmf	1.11 s	0.057	0.020	92.8%
pdal filters.smrf	1.73 s	0.035	0.020	94.0%
pdal filters.csf	6.49 s	0.121	0.090	88.1%

Two things follow. Our PMF is correct — it ties PDAL's own implementation at twice the
speed, so this is not a bug fix. And SMRF is 37% lower RMSE with 1.3 points more
coverage, which is a real quality gain we do not currently have.

Ground quality bounds everything downstream: canopy height is height above ground, so it
propagates into the CHM, the tree heights ITD derives from it, and the canopy fuels built
from those.

Why implement rather than depend on PDAL

Adopting PDAL for this was investigated and rejected on measurement:

• 9-13× slower end to end for the chunked CHM (hag_nn), 25-37× with hag_delaunay.
• 709 MB / 181 conda packages, including a second GDAL beside the one rasterio already
  bundles. There is no pip route — pdal on PyPI is an sdist needing libpdal present.
• Chunk + buffer + stitch never becomes exact: seam cells differing from an unchunked
  run go 16.08% (no buffer) → 0.36% (15 m) → 0.03% (140 m), with a 0.39 m max still
  present at 140 m. Our blocked implementation is bit-identical to unblocked by
  construction, because its halo is on rasters rather than points.

SMRF is a published algorithm, so we can have the accuracy without any of that.

Validation

services/griddle/tests/local/test_chm_vs_pdal.py already does exactly this job for PMF —
PDAL filters, our code rasterises in both arms, so the comparison isolates the algorithm
from the gridding convention. Add an SMRF case to it.

Target: match filters.smrf as closely as _pmf matches filters.pmf today — median
0.00 m, p95 0.04 m, 98.3% of cells within 10 cm.

The tests are local-only (local marker, skipped without PDAL), so CI is unaffected.

Notes

• Exposing the choice of ground algorithm on the create request is Expose ground-source and ground-derivation controls on the point-cloud CHM request #484; this issue is the
  algorithm itself, and Expose ground-source and ground-derivation controls on the point-cloud CHM request #484 is what would surface it.
• filters.csf was measured and is worse on this data — not worth pursuing.

[amarcozzi]

Ordering note before anyone picks this up: the validation harness this issue depends on is not on main yet.

This issue says services/griddle/tests/local/test_chm_vs_pdal.py "already does exactly this job for PMF — add an SMRF case to it." That file does not exist on main. It is 626 new lines on lakitu-parquet-write-path (#492), along with chm_blocks.py and 718 changed lines of chm_point_cloud.py.

So starting SMRF today means one of:

• branch off Store point clouds as spatially partitioned Parquet #492 and stack on an unmerged ~1,800-line change, or
• implement against main's chm_point_cloud.py, which has no PDAL comparison to validate with — and this issue is entirely a claim about measured accuracy, so an unvalidated implementation would not close it.

Neither is worth it. Blocked on #492 merging, same as the fixture restore.

Also worth flagging while this is open: #503 was filed today for a density-dependent height bias in the same handler, and it interacts with this one. The derived-ground figures there (+0.55 m bias, 7.1% of cells over 1 m) are measured with the current PMF, so SMRF will move them. The classified-ground figures (+0.24 m) it will not touch, because it does not change which returns are ground on that path. Whoever does this should re-measure #503's derived rows afterwards rather than treating the two as independent.