flow_length_mfd: no memory guard on H*W working arrays

[brendancol]

flow_length_mfd in xrspatial/hydro/flow_length_mfd.py allocates several full-grid working arrays in its eager numpy and cupy backends with no upfront budget check. A 50000x50000 raster asks for about 73 GB of host RAM before anything errors out.

Same shape as #1332 for flow_length_d8, and the broader hydro audit started with #1318/#1319 for flow_accumulation_d8. See also #1341 (stream_link_mfd) for the MFD overhead pattern. Eager backends check, dask backends skip since per-tile allocations are bounded by chunk size.

Allocations

_flow_length_mfd_downstream_cpu and _flow_length_mfd_upstream_cpu (called by both numpy and cupy backends) allocate:

Array	dtype	bytes/px
in_degree	int32	4
valid	int8	1
flow_len	float64	8
order_r	int64	8
order_c	int64	8
Total		29

The numpy public dispatch also performs frac = data.astype(np.float64), an explicit 64 B/px copy of the (8, H, W) MFD fractions input that lives on top of the working set. Total numpy peak: 29 + 64 = 93 B/px.

For the cupy backend, _flow_length_mfd_cupy calls .get().astype(np.float64) then runs the CPU kernel, so the host pays the full 93 B/px during the call. On the device side the input (8, H, W) float64 already exists (64 B/px) and the output float64 result lives at 8 B/px - call it 100 B/px GPU as a conservative budget covering both copies plus headroom.

Worked example

50000x50000 = 2.5e9 pixels. CPU peak working set:

2.5e9 * 29 = 73 GB (kernel locals only)
2.5e9 * 93 = 232 GB (with the up-front astype copy)

Allocated before any sanity check runs.

Fix

Mirror PR #1332: _check_memory and _check_gpu_memory helpers, wired into the numpy and cupy dispatch in flow_length_mfd(). Leave dask alone. Add tests for oversize rejection, valid pass-through, dask bypass, and error message content.

One fix per PR per the security-sweep policy.