Add soil module (SoilGrids 2.0) with NEH 630 Ch7 hydrologic soil group

[rehsani]

Summary

• New `floodpath.soil` module: ISRIC SoilGrids 2.0 (250 m global, COG via `/vsicurl/` + `WarpedVRT` from native Goode Homolosine to WGS84). Fetches sand/silt/clay (% by mass) at user-selectable depths and derives the NEH Part 630 Chapter 7 hydrologic soil group (A/B/C/D).
• HSG is the SOIL half of the SCS-CN runoff parameterisation. Combined with `floodpath.landuse` (already in place) it lets a future `hydrology.runoff` module compute Curve Numbers.

Module surface

Item	What
`get_soilgrids_texture(lat, lon, buffer_deg, depths)`	Fetch sand/silt/clay as a depth-weighted `TextureGrid`. Default depths = (0-5cm, 5-15cm, 15-30cm), the closest proxy for NEH 630 Ch7's 0-50 cm reference.
`TextureGrid`	sand/silt/clay rasters + georef + `closure_residual()` diagnostic. Nodata (-32768) is masked to NaN.
`usda_texture_class(sand, silt, clay)`	Full USDA Soil Survey Manual 1993 texture triangle, 12 canonical classes. Scalar or vectorised.
`texture_to_hsg(TextureGrid)`	`USDA_TEXTURE_TO_HSG` lookup (NEH 630 Ch7 §630.0701 textural shortcut) -> `HSGGrid`. Custom mappings supported.
`HSGGrid`	uint8 raster (1=A, 2=B, 3=C, 4=D, 0=nodata) + `class_counts()` / `fraction()` / `dominant_class()` / `as_letter_grid()`

Performance

SoilGrids reads were the bottleneck — 9 sequential `/vsicurl/` fetches for 3 vars × 3 depths took ~525 s. Parallelised via `ThreadPoolExecutor` (max 9 workers): ~73 s for the same workload (~7× speedup). ISRIC tolerates ~10 concurrent connections per client.

Test plan

• `pytest -m "not integration"` — 211 passed (154 prior + 57 new soil tests; 16 deselected integration)
• `pytest -m integration tests/soil/` — 3 passed (~169 s): live ISRIC fetch in Djibouti at default depths, single-depth fetch, fixture regression for the Robit Bata patch
• All 12 USDA texture classes round-trip from textbook centroids; all four HSG groups (A/B/C/D) reachable
• Robit Bata patch (Vertisol-rich Ethiopian highlands) classifies as 99.9% HSG D — geologically correct (cracking black clays, ~54% mean clay)
• End-to-end smoke test on Robit Bata: 14 stages all succeed, including the new HSG plot at stage 12
• Committed fixture is 8.4 KB (`tests/fixtures/robit_bata_soil.tif`, 33×33 multiband float32) — well under the per-fixture budget

Notes

• Default texture-shortcut HSG assignment assumes "deep" soils with no shallow restrictive layer or high water table. Users with site-calibrated Ksat can supply a custom `mapping` to `texture_to_hsg` to bypass the textural shortcut.
• `TextureGrid.closure_residual()` reports per-cell `|sand+silt+clay - 100|`. SoilGrids regressions are independent for each variable so closure is not exact (typical ~0.02% mean residual on the Robit Bata fixture).
• Single white nodata pixel in the smoke-test HSG plot is the Goode → WGS84 reproject corner artefact and is expected.