Add routing module: steady-state runoff accumulation + peak discharge (1/2)

[rehsani]

Summary

PR 1 of 2 for the routing milestone. Closes the hydrologic half of routing — turns SCS-CN per-cell runoff (mm) into accumulated upstream volume (m³) and peak discharge (m³/s) along the flow direction grid. The follow-up PR adds Manning normal-depth at stream cells to close the hydraulic half (discharge → water level → HAND-driven inundation).

Module surface

Item	What
`accumulate_runoff(runoff, flow_grid)`	Reproject Q (10 m → 30 m via area-weighted average) → mm × cell-area → m³ → `pyflwdir.accuflux` downstream
`peak_discharge(accumulated, duration_s)`	Q_peak = V_acc / T per cell
`cell_areas_m2(transform, shape, crs)`	Handles WGS84 cos(lat) shrinkage; works for projected CRSes too
`AccumulatedRunoffGrid`, `DischargeGrid`	Carry method/precip-source attribution + helper stats

Why steady-state, why two PRs

• Steady-state (V/T) is appropriate for v0.2 — matches the project's "single event snapshot" model and gives a deterministic answer without time stepping. Time-resolved hydrographs (unit hydrograph or kinematic wave) are a v0.3 ambition.
• Two PRs keeps each diff reviewable. PR1 (this one) gives a discharge field you can plot and validate; PR2 adds the hydraulic closure (Manning normal-depth → water level at stream cells → integration into HAND inundation).

Resolution alignment

`accumulate_runoff` reprojects the runoff raster (WorldCover 10 m grid) onto the DEM 30 m grid using `Resampling.average` — the area-weighted mean is the correct aggregation for a depth quantity. Each coarse cell's volume is then `mean_depth × cell_area`, where cell area handles WGS84 longitude convergence.

Test plan

• `pytest -m "not integration"` — 294 passed (264 prior + 30 new routing tests; 16 deselected integration)
• `cell_areas_m2` sanity: equator vs 60° N (cos(60°)=0.5 → half area), southern vs northern hemisphere parity
• Peak-discharge arithmetic: V/T pinned at 1.0 m³/s for 21,600 m³ over 6 hr; halving T doubles Q; negative T raises
• Robit Bata pinned values: outlet V_acc = 1.43 Mm³ (matches independent water-balance estimate), outlet peak Q = 66.2 m³/s for 100 mm × 6 hr event
• Outlet of accumulated runoff matches the cell with maximum flow accumulation (same flow direction grid drives both)
• End-to-end smoke test on Robit Bata: 17 stages all succeed, including new stage 15 (two-panel V_acc + Q_peak with stream overlay)

Notes

• The bbox is square-cropped, not a closed basin — most cells drain off the edges. Outlet accumulation captures only the basin upstream of the bbox's true outlet (~31% of the patch in Robit Bata's case). This is correct.
• Steady-state assumption breaks down for large basins where peak attenuation matters. Document this as a known limitation; the kinematic-wave path is left for v0.3.