Add viewshed tests for 1x1, all-NaN, and Inf-terrain edge cases

[brendancol]

Closes #3425

Test-only change from the test-coverage sweep. Three input shapes that nothing in test_viewshed.py exercised, all already behaving correctly:

• test_viewshed_single_pixel: a 1x1 raster returns [[180.0]] (the degenerate-axis resolution guard from viewshed() returns wrong results for single-row or single-column rasters (divide-by-zero in resolution) #2744 keeps it from dividing by zero)
• test_viewshed_inf_terrain_cell: an inf elevation cell comes back visible and doesn't leak NaN into the rest of the grid
• test_viewshed_all_nan_input: an all-NaN raster gives observer=180 and INVISIBLE everywhere else

Each is parametrized over numpy and dask Tier B (the CPU-backed backends). The cupy/RTX path is a separate implementation and is out of scope here.

No source change. I probed all three by hand first to confirm the behavior was already correct, so these pin existing behavior rather than fixing a bug.

Test plan:

• pytest xrspatial/tests/test_viewshed.py — 120 passed, 1 skipped (no-RTX fallback)
• the 6 new test cases run and pass on a CUDA+RTX host
• flake8 clean on the test file

[brendancol]

PR Review: viewshed edge-case test coverage

Test-only change, three new tests parametrized over numpy and dask Tier B. I read all three in full and ran the suite. No blockers, no suggestions.

Blockers

None.

Suggestions

None.

Nits

• test_viewshed_all_nan_input hardcodes the observer's flat index as 2 * 5 + 2. It's tied to the fixed 5x5 grid and the inline comment explains it, so this is fine as-is; just noting it for anyone who later changes the grid size.

What looks good

• Each test checks against known values (observer cell = 180, INVISIBLE elsewhere), not just "runs without crashing".
• NaN checks use np.isnan, and the inf test explicitly asserts no NaN leaks into the rest of the grid.
• Backend scope (numpy + dask Tier B, cupy/RTX out) matches the existing NaN tests in this file and is called out in each docstring.
• 1x1 dask uses a (1,1) chunk, which exercises the degenerate-axis resolution guard on the dask path too.

Checklist

• Algorithm matches reference/paper (n/a — test-only)
• All implemented backends produce consistent results (numpy + dask asserted; cupy/RTX scoped out)
• NaN handling is correct
• Edge cases are covered by tests (that's the PR)
• Dask chunk boundaries handled correctly
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (not needed — tests)
• README feature matrix updated (n/a)
• Docstrings present and accurate