Add runoff module with SCS Curve Number derivation

[rehsani]

Summary

• New `floodpath.runoff` module: combines ESA WorldCover landuse with the NEH 630 Ch7 hydrologic soil group into a per-cell SCS Curve Number raster — the missing link between the soil + landuse data layers and the SCS-CN runoff equation.
• Default `WORLDCOVER_NEH_CN` lookup uses NEH 630 Ch9 Table 9-1 typical values (AMC II) with documented WorldCover-to-NEH cover-type mappings; users can override per-class CN values for site-calibrated work.

What this enables

`CurveNumberGrid.potential_retention_mm()` returns `S = 25400/CN - 254` ready to feed into `Q = (P - 0.2S)² / (P + 0.8S)`. Once a rainfall source lands (next PR), the SCS-CN equation itself is a 5-line addition.

Module surface

Item	What
`compute_curve_number(landuse, hsg, mapping=None, fallback=70)`	Pure transformation. HSG (~250 m) is upsampled to landuse grid (10 m) via nearest-neighbour reproject.
`CurveNumberGrid`	uint8 CN raster + `stats()` that ignore nodata + `potential_retention_mm()`
`WORLDCOVER_NEH_CN`	Default 11×4 lookup (NEH 630 Ch9 Table 9-1)

Test plan

• `pytest -m "not integration"` — 240 passed (211 prior + 29 new runoff tests; 16 deselected integration)
• Table invariants: every WorldCover class has all four HSG letters, CN ∈ [30,100], CN monotone-increasing across A→D for every cover type
• Pinned NEH-iconic values: cropland × D = 89, woods × A = 30, water = 100, snow/ice = 98
• S formula round-trip: CN=100 → S=0, CN=77 → S=75.87 mm, CN=50 → S=254 mm exactly
• Pinned Robit Bata CN derivation against committed fixtures: mean CN = 87.72, dominant value 89 (cropland on D), 809,271/810,000 cells classified
• End-to-end smoke test on Robit Bata: 15 stages all succeed, including new CN plot at stage 13

Notes

• WorldCover-to-NEH cover-type mappings (e.g. `built_up` → "Residential 1/4 acre") are documented in `runoff/constants.py` with a note that users with dense urban patches should override (e.g. with the NEH "Commercial" curve A=89/B=92/C=94/D=95).
• 729 nodata cells in the Robit Bata CN raster come from the single SoilGrids HSG nodata pixel projected onto the 27× finer WorldCover grid — expected behaviour.