Compute dask source once in cumulative_viewshed (#3185)

[brendancol]

Closes #3185

What changed

• cumulative_viewshed called viewshed() once per observer. On the dask path with no max_distance, every call ran raster.data.compute() on the same source, so the source was materialised once per observer.
• When every observer takes the full-grid path (no function-level or per-observer max_distance), the source is now computed once up front and the observers run against the in-memory raster. The accumulated count is re-wrapped as dask so the output backend still matches the dask input.
• Observers that set a max_distance keep the dask windowing path, which loads only each observer's window.

Backend coverage

• numpy: unchanged.
• dask+numpy and dask+cupy: source computed once instead of N times on the no-max_distance path; per-observer max_distance behaviour unchanged.
• cupy: unchanged.

Test plan

• test_dask_source_computed_once asserts a single source compute for four observers and a dask-backed output.
• test_dask_per_observer_max_distance_stays_lazy asserts the windowed path stays dask and matches numpy.
• Existing test_dask_matches_numpy parity check still passes.
• Full test_visibility.py suite: 27 passed.

[brendancol]

PR Review: Compute dask source once in cumulative_viewshed

(Posted as a comment via REST; gh pr review/gh pr comment GraphQL path returned HTTP 401.)

Blockers

None.

Suggestions

• The da.from_array(count, chunks=_input_dask_chunks) wrap at the end builds a dask array whose data is already a concrete numpy buffer. Intentional and correct (it restores the dask output type without re-running the sweep), but the comment only says "Restore the dask-backed output type the dask input would have given." A half-line noting the count is already concrete, so nothing lazy is recovered, would save the next reader a double-take. Not blocking.

Nits

• The new guard not any('max_distance' in obs for obs in observers) walks the observer list once before the main loop. Fine for normal observer counts.

What looks good

• Surgical change matching the issue: compute the dask source once when every observer takes the full-grid path, otherwise keep per-observer dask windowing.
• Output backend preserved: a dask input still returns a dask-backed result, output chunks match the input (verified).
• Caller's raster left alone; raster.copy() does not compute or mutate the input (verified).
• Correctness holds: Tier B already computed to numpy and ran _viewshed_cpu; existing test_dask_matches_numpy still passes.
• Tests: test_dask_source_computed_once asserts one source compute for four observers and a dask output; test_dask_per_observer_max_distance_stays_lazy asserts the windowed path stays dask with numpy parity.
• Verified locally: source computes dropped from N to 1; computing the result adds zero further source computes.

Checklist

• Algorithm matches reference: N/A (scheduling change only)
• Backends consistent: yes
• NaN handling: unchanged
• Edge cases tested: yes (no max_distance, per-observer max_distance, parity)
• Dask chunk boundaries: yes (output chunks match input)
• No premature materialization or unnecessary copies: improved (removes N-1 source materializations)
• Benchmark: none; not needed for a scheduling fix
• README feature matrix: N/A
• Docstrings: accurate (no signature change)