Guard hand_dinf() against unbounded memory allocations (#1344)

[brendancol]

Closes #1344.

Same shape as #1326 (hand_d8) and #1340 (hand_mfd). The numpy and cupy backends of hand_dinf both run _hand_dinf_cpu, which allocates seven full-grid working arrays totalling 38 bytes/pixel. A 50000x50000 raster would ask for ~95 GB before any sanity check ran.

Allocations

_hand_dinf_cpu allocates:

Array	dtype	bytes/px
in_degree	int32	4
valid	int8	1
is_stream	int8	1
drain_elev	float64	8
hand_out	float64	8
order_r	int64	8
order_c	int64	8
Total		38

The cupy path copies its three inputs to host via .get() then runs the same CPU kernel, so it also pays 38 B/px on the host plus the input copies. A conservative 32 B/px applies on the device side for symmetry with #1326 and #1340.

Fix

Adds _check_memory and _check_gpu_memory helpers and wires them into the numpy and cupy branches of hand_dinf(). Dask paths are untouched since per-tile allocations are bounded by chunk size.

Tests

Five new tests cover:

• numpy oversize raster raises MemoryError
• numpy normal raster passes the guard
• dask path skips the guard even when memory is reported as 1 byte
• error message includes grid dimensions and a "use dask" hint
• cupy oversize raster raises MemoryError (CUDA-only)

Full hydro suite runs 635 passed, 1 failed (test_stream_link_dask_temp_cleanup, a known flake unrelated to this change).