Validate viewshed max_distance (reject negative / non-finite)

[brendancol]

Closes #2855

viewshed() didn't validate max_distance. Negative, NaN, and infinite
values fell through to the windowing math and raised internal errors that
never mentioned the argument (for example "zero-size array to reduction
operation minimum" for -1.0 and "cannot convert float NaN to integer" for
NaN).

What changed:

• Validate max_distance at the public viewshed() entry point, before
  backend dispatch, and raise ValueError("max_distance must be a finite number >= 0, ...") for negative or non-finite values.
• Document the requirement on the max_distance parameter.

Backend coverage: the check runs before dispatch, so it applies to all
four backends (numpy, cupy, dask+numpy, dask+cupy).

Test plan:

• New test asserts the clear ValueError for -1.0, -0.5, NaN, +inf, -inf.
• New test confirms finite values (0.0, 3.0) still work on numpy and dask+numpy.
• Full test_viewshed.py passes (72 passed, 1 xfailed).

Skipped the user-guide-notebook and README-matrix steps: this is a pure
input-validation bug fix with no new public API and no change to backend
support.

[brendancol]

PR Review: Validate viewshed max_distance (reject negative / non-finite)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• xrspatial/viewshed.py:1721 and 1729: there are now two back-to-back if max_distance is not None: blocks. You could fold the validation into the existing block, but keeping them separate reads fine too (validation first, windowing dispatch under its own comment). Optional.

What looks good

• Validation sits at the public entry point before backend dispatch, so all four backends (numpy, cupy, dask+numpy, dask+cupy) are covered by one check. The test docstring spells out that reasoning.
• The error message names the argument and echoes the bad value, which is the point of the fix.
• The try/except (TypeError, ValueError) also turns a non-numeric max_distance into the same clear message instead of a raw np.isfinite TypeError.
• Tests cover -1.0, -0.5, NaN, +inf, -inf and confirm 0.0 and 3.0 still work on numpy and dask.
• Docstring updated to state the contract.

Checklist

• Validation runs before dispatch, so all backends behave the same
• NaN / inf / negative handled with a clear ValueError
• Edge cases (0.0, +inf, -inf, non-numeric) covered
• No new materialization or copies
• No new public API, so README matrix and user guide are not needed
• Docstring present and accurate

[brendancol]

Follow-up review (after 9104c38)

The one nit from the first pass is resolved: the validation and the windowed dispatch now share a single if max_distance is not None: branch (xrspatial/viewshed.py:1721). Behavior is unchanged and the 24 max_distance tests still pass.

No remaining Blockers, Suggestions, or Nits.