Validate hotspots() kernel like apply()/focal_stats()

[brendancol]

Summary

hotspots() skipped the kernel validation that apply() and focal_stats() run via custom_kernel(), so bad kernels produced confusing errors or wrong behavior:

• Run custom_kernel(kernel) before the memory check. None and list-of-list kernels now raise a clear ValueError instead of AttributeError on kernel.shape, and even-dimensioned kernels are rejected for improper shape (previously they silently succeeded).
• Add a zero-sum kernel guard. hotspots() is the only focal function that normalizes by kernel.sum(), so a zero-sum kernel previously divided by zero.

Backend coverage

The validation runs before backend dispatch, so it applies to numpy, cupy, dask+numpy, and dask+cupy.

Test plan

• New regression tests for None, list-of-list, even-dim, and zero-sum kernels
• Valid-kernel happy path still produces the expected classification
• Full test_focal.py suite passes (155 passed)

Closes #2771

[brendancol]

PR Review: Validate hotspots() kernel like apply()/focal_stats()

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• focal.py:1509 -- the guard catches exactly-zero sums. A mixed-sign kernel that sums to roughly zero but not exactly zero (floating-point near-cancellation) still slips through and produces a near-infinite normalization. This matches focal_stats() and is within the issue scope, so leaving it is fine.

What looks good

• Validation runs before backend dispatch, so numpy, cupy, dask+numpy, and dask+cupy all reject bad kernels the same way. The 3D path recurses through hotspots() per band, so each band validates too.
• custom_kernel() runs before kernel.sum(), so the sum call always has an ndarray. No AttributeError risk.
• Error messages name the function and explain the failure.
• Tests cover None, list-of-list, even-dim, and zero-sum kernels plus a valid happy path. Full test_focal.py passes (155).

Checklist

• Algorithm unchanged; only input validation added
• All backends get consistent validation (runs before dispatch)
• NaN handling unchanged
• Edge cases covered by tests
• Dask path unaffected (validation is pre-dispatch)
• No premature materialization
• Benchmark not needed (validation-only change)
• README feature matrix not applicable (no new function)
• Docstrings present (existing hotspots docstring unchanged)