Add precip module + SCS-CN equation: end-to-end rainfall->runoff pipeline

[rehsani]

Summary

• New `floodpath.precip` module: `PrecipGrid` dataclass + `uniform_precip` helper. Symmetric with `landuse`/`soil` so future fetchers (ERA5 / IMERG / CHIRPS) slot in cleanly.
• `floodpath.runoff.apply_scs_cn(cn, precip)` implements the SCS-CN equation `Q = (P - 0.2S)² / (P + 0.8S)` with the initial-abstraction threshold (P ≤ 0.2S → Q=0). Closes the rainfall-runoff loop for the project — the pipeline is now end-to-end without waiting on a precip data-source decision.

Why now (per discussion)

Building a uniform-precip input first lets us validate the SCS-CN math and the whole chain (landuse + soil → CN → Q) before tackling the precip data-source choice. When ERA5/IMERG/CHIRPS lands later, it returns a `PrecipGrid` and feeds the same `apply_scs_cn` — no consumer change.

Resolution alignment (the question that motivated this PR)

Each layer keeps its native resolution; `compute_curve_number` already upsamples HSG (~250 m) to the landuse grid (10 m) via nearest-neighbour. The eventual common grid for routing will be DEM 30 m; that reprojection step properly belongs to the routing module (next milestone after a real precip source).

Test plan

• `pytest -m "not integration"` — 264 passed (240 prior + 24 new; 16 deselected integration)
• SCS-CN equation pinned at canonical CN/P values: CN=89, P=100 → Q=70.21; CN=77, P=100 → Q=44.78; CN=100 → Q=P (water passes everything through)
• Asymptotic limit: as P → ∞, Q → P − 1.2·S (verified for CN=80 at P=10000)
• Monotonicity: higher CN → higher Q for the same rainfall
• Nodata propagates: CN=0 cells produce NaN in Q (NOT 0), so unmodelled and zero-runoff cells stay distinguishable
• Custom `initial_abstraction_ratio=0.05` (modern recalibration) gives more runoff than the NEH default 0.2
• End-to-end smoke test on Robit Bata: 16 stages, including new stage 14 (uniform 100 mm → SCS-CN runoff). Mean Q = 67.5 mm, C = Q/P = 0.674 — physically meaningful for clay-rich cropland.

Notes

• Default `lambda = 0.2` per NEH 630 Ch10. Pass `initial_abstraction_ratio=0.05` for the newer recalibration (Hawkins et al. 2002).
• `RunoffGrid.precip_source` records the source string (`"uniform"` here) so downstream consumers can attribute results when we eventually mix precipitation sources.
• Smoke test exposes `--uniform-precip-mm` (default 100 mm).