Add Manning normal-depth + rainfall-driven HAND flood (2/2)

[rehsani]

Summary

PR 2 of 2 for the routing milestone. Closes the hydraulic half — turns the per-cell peak discharge from PR1 into per-stream-cell water levels via Manning's wide-channel normal-depth equation, then broadcasts those upstream along the flow direction grid to drive HAND-based inundation. Combined with PR1, the project now produces a true rainfall-driven flood depth raster from synthetic uniform rainfall (or any future `PrecipGrid`).

Pipeline closed

```
rainfall (uniform or any source)
↓ SCS-CN equation
runoff Q (mm)
↓ accumulate_runoff (PR1)
V_acc (m³)
↓ peak_discharge (PR1)
Q_peak (m³/s)
↓ compute_water_level — NEW
h at stream cells (m)
↓ compute_rainfall_inundation — NEW
flood depth (m) per cell
```

Module surface

Item	What
`slope_along_flow(flow_grid, dem)`	Per-cell slope (m/m) along D8 with metric distances; min_slope floor for Manning sqrt safety
`leopold_maddock_width(Q, a=2.5, b=0.5)`	Global natural-channel hydraulic-geometry default
`manning_normal_depth(Q, w, S, n)`	Closed-form wide-channel solution `h = ((Q·n)/(w·√S))^(3/5)`
`compute_water_level(discharge, roughness, flow_grid, streams, dem)`	Convenience: ties everything together; output is `WaterLevelGrid` with h at stream cells, NaN elsewhere
`compute_rainfall_inundation(water_level, hand, flow_grid, streams)`	Broadcasts per-stream h upstream via `pyflwdir.basins`, applies `depth = max(W - HAND, 0)`
`WaterLevelGrid`, `RainfallInundationDepth`	Carry full provenance (precip_source, duration_s, method strings)

Test plan

• `pytest -m "not integration"` — 323 passed (294 prior + 29 new; 16 deselected integration)
• Slope helper: clamped to floor on flat DEM; positive on inclined; mountain-catchment range on Robit Bata (median ~5%)
• Manning: pinned hand-computed h=1.150 m for Q=10, w=10, S=0.001, n=0.04; monotonic in n (rougher → deeper) and S (steeper → shallower); zero/negative inputs handled
• End-to-end on Robit Bata fixtures: outlet stream h = 10.07 m (matches independent computation), flooded fraction = 9.5% under 100 mm × 6 hr — physically plausible for a small clay-rich Ethiopian catchment
• All non-stream cells in `WaterLevelGrid` are NaN by construction; stream cells are finite
• `compute_rainfall_inundation` raises `ValueError` on shape mismatch
• Smoke test stage 16 produces a hillshade + log-flood-depth + stream-overlay map showing the dendritic flood pattern

Notes

• Defaults follow Andreadis et al. 2013 (LM coefficients) and standard NEH min-depth/min-slope conventions; all overridable.
• `WaterLevelGrid` masks non-stream cells to NaN to make the field semantically clean — only channels carry water level. `RainfallInundationDepth` then broadcasts those values upstream to every draining cell.
• Out-of-basin cells (ridge cells in the square-cropped bbox that drain off the edge before hitting any stream) stay dry, which is the correct behaviour.
• Time-resolved (kinematic-wave or unit-hydrograph) routing remains a v0.3 ambition; this PR uses the steady-state V/T peak-discharge approximation appropriate for small basins and intense storms.