flow_path_mfd: no memory guard on H*W working arrays

[brendancol]

Summary

flow_path_mfd() allocates several (8, H, W) and (H, W) float64 working arrays in the numpy and cupy backends without any pre-allocation memory check. Large rasters can exhaust host or device RAM and abort the process before any error message reaches the caller.

This continues the hydro memory-guard series (#1318, #1319, ..., #1351, #1355, #1357).

Affected backends

• numpy (flow_path_mfd -> _flow_path_mfd_cpu via the public dispatch)
• cupy (_flow_path_mfd_cupy does .get() + .astype(float64) host-side)

The dask and dask+cupy paths process tiles and are bounded by chunk size, so they skip the guard.

Working-memory accounting

CPU peak working set (B/px):

• data.astype(np.float64) copy of (8, H, W) fractions: 64
• np.asarray(sp_data, dtype=np.float64) copy of start_points: 8
• out (H, W) float64 written by _flow_path_mfd_cpu: 8

Total: ~80 B/px.

GPU peak working set (B/px), _flow_path_mfd_cupy:

• Host-side .get() of (8, H, W) fractions: 64
• Host-side .astype(np.float64) copy: up to 64 (skipped when dtype already matches)
• Host-side output (H, W) float64: 8
• Device-side output (H, W) float64: 8

Conservative GPU budget: ~88 B/px.

Worked example

A 30000 x 30000 fractions raster:

• numpy: 30000 * 30000 * 80 = 72 GB peak working memory.
• cupy: 30000 * 30000 * 88 = ~79 GB host residency during the host-side trace.

Both far exceed typical workstation RAM, and no MemoryError is raised before allocation.

Proposed fix

Match the pattern from #1351 (flow_length_mfd) and #1337 (stream_link_mfd):

• Per-module _BYTES_PER_PIXEL and _GPU_BYTES_PER_PIXEL constants with a comment breaking down the count.
• _available_memory_bytes() reads /proc/meminfo then falls back to psutil, then a 2 GiB default.
• _available_gpu_memory_bytes() queries cupy.cuda.runtime.memGetInfo, returns 0 if unavailable.
• _check_memory(H, W) raises MemoryError if projected use exceeds 50% of available host RAM.
• _check_gpu_memory(H, W) mirrors the host check on the GPU; skips silently when CUDA is unavailable.
• Wire both checks into the public flow_path_mfd dispatch before the eager allocations. Dask and dask+cupy paths skip the guard.

Tests

• numpy raster raises MemoryError when projected RAM exceeds budget (mock _available_memory_bytes).
• numpy raster of normal size succeeds.
• dask path bypasses the guard even when memory is mocked low.
• error message mentions grid dimensions and points to dask.
• cupy raster raises MemoryError when projected GPU RAM exceeds budget (skipped without CUDA).