Fix viewshed crash on NaN cells and mark them INVISIBLE (#2857)

[brendancol]

Closes #2857

What changed

• _init_event_list skips NaN cells so they no longer emit ENTERING/CENTER/EXITING events. This removes the ValueError: node not found that hit when a NaN sat on the observer row to the right of the observer (its EXITING event tried to delete a node that was never inserted).
• _init_event_list now returns the event count, and _viewshed_cpu trims the unused trailing rows of the pre-allocated event list before sorting. Without the trim, the leftover all-zero rows sort as CENTER events at cell (0, 0) and would mark it visible.
• NaN cells keep their INVISIBLE (-1) fill value. That is the sentinel downstream consumers already test for (visibility.py and experimental/min_observable_height.py both treat != INVISIBLE as visible), so a NaN cell is never counted as visible.

Backends

The fix lives in the CPU sweep, which backs the numpy backend and the dask Tier B path. The cupy/RTX viewshed is a separate implementation and is out of scope here.

Test plan

• Converted the xfail NaN test into passing tests.
• NaN at every position around a 5x5 grid, including the observer row to the right.
• Guard that skipped NaN cells don't leak visibility to cell (0, 0).
• numpy and dask+numpy backends.
• test_viewshed.py, test_visibility.py, test_min_observable_height.py all green.

No new public API, so the user-guide notebook and README feature-matrix steps are skipped. No docs change needed; the viewshed reference entry and signature are unchanged.

[brendancol]

PR Review: Fix viewshed crash on NaN cells and mark them INVISIBLE (#2857)

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• The new if np.isnan(inrast[1][j]): continue reads inrast[1][j], the same value the line above stores into e[E_ELEV_1] (viewshed.py:1165). Reusing e[E_ELEV_1] for the guard would drop the second index, but reading inrast[1][j] directly next to the comment is clearer. Wash. Leave as-is.

What looks good

• Root cause is right. NaN cells emitted EXITING events with no matching insert (the pre-insert loop at viewshed.py:1388 guards on not np.isnan), so _delete_from_tree raised. Skipping event creation for NaN cells removes the asymmetry.
• The trailing-row trim (event_list = event_list[:count_event]) is load-bearing, not cosmetic. Without it the leftover all-zero rows sort as CENTER events at (0, 0) and would mark that cell visible. test_viewshed_nan_does_not_leak_visibility_to_origin covers that path.
• Sentinel choice is correct. NaN cells stay at the INVISIBLE fill value, which is what visibility.py:287 and experimental/min_observable_height.py:171 test against. A NaN output would have read as visible; INVISIBLE does not.
• Verified two more edge cases locally: a NaN observer cell and an all-NaN raster both return cleanly (observer 180, rest INVISIBLE) with no crash.
• The dask Tier B path runs through _viewshed_cpu, so it picks up the fix. The dask+numpy test confirms it.
• The Tier C out-of-core distance sweep already skips NaN in _sweep_ring (elev != elev), so it was never affected and the fix correctly leaves it alone.

Checklist

• Algorithm matches reference (GRASS r.viewshed port; NaN cells are NODATA and excluded)
• Implemented backends consistent (numpy + dask Tier B share the CPU sweep; cupy/RTX is a separate path, out of scope)
• NaN handling correct
• Edge cases covered by tests
• Dask chunk boundaries handled (Tier B computes to one in-memory array)
• No premature materialization or unnecessary copies
• Benchmark not needed (bug fix, no perf change)
• README feature matrix N/A (no new function)
• Docstrings accurate (no API change)