Guard basin_d8() against unbounded memory allocations (#1357)

[brendancol]

Closes #1357.

Summary

basin_d8() on the numpy and cupy backends allocated H*W working arrays without checking available memory:

• numpy: fd cast (8 B), labels (8 B), state (1 B), plus path_r/path_c int64 inside _watershed_cpu (8 + 8 B). ~33 bytes/pixel.
• cupy: flow_dir_f64, labels, state, and the final cp.where result. ~28 bytes/pixel on the device.

A 50000x50000 input on CPU would request ~83 GB before erroring out.

Fix

Add module-local _BYTES_PER_PIXEL / _GPU_BYTES_PER_PIXEL, _available_memory_bytes / _available_gpu_memory_bytes, _check_memory / _check_gpu_memory helpers wired into the numpy and cupy dispatch branches before allocations. 50% threshold matches the watershed_d8 / flow_accumulation_d8 / hand_d8 series.

The dask paths delegate to the watershed dask infrastructure and stay bounded per-tile by the user's chunk size, so they skip the guard.

Tests

5 new tests in TestMemoryGuard:

• numpy raises MemoryError when _available_memory_bytes is patched low
• numpy normal-sized input succeeds against real memory
• dask path skips the guard and computes
• error message includes the grid dimensions and a dask hint
• cupy raises MemoryError when _available_gpu_memory_bytes is patched low (skipped without CUDA)

Test plan

• pytest xrspatial/hydro/tests/test_basin_d8.py -- 22 passed
• pytest xrspatial/hydro/tests/ -- 644 passed