viewshed: consistent output name and dtype across backends

[brendancol]

Closes #2743

viewshed() returned a DataArray whose .name and dtype depended on which backend ran:

• .name: numpy left it None, the RTX path hardcoded "viewshed", and the dask paths leaked the internal dask graph token (built without an explicit name=).
• dtype: float32 on the GPU/RTX exact path, float64 everywhere else.

Changes:

• Add a name= parameter (default 'viewshed'), threaded through every backend helper and set on the output. Matches the convention in slope, curvature, and hillshade.
• Force float64 output on the GPU and dask Tier B paths so dtype no longer depends on the backend (the RT kernel still works in float32 internally).

attrs, coords, and dims were already preserved on every backend; this PR does not change them.

Backend coverage: numpy, cupy (RTX), dask+numpy, dask+cupy. Verified live on a host with CUDA + RTX.

Test plan:

• test_viewshed_output_name_and_dtype_consistent — name == 'viewshed' and dtype == float64 across all four backends, with and without max_distance
• test_viewshed_custom_name — a user-supplied name is honoured on every backend
• Full test_viewshed.py suite (37 passed)
• test_visibility.py, test_min_observable_height.py, test_gpu_rtx_memory.py, test_validation.py (120 passed)

[brendancol]

PR Review: viewshed consistent output name and dtype across backends

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None.

Nits (optional improvements)

• The docstring lists the name type as str, but the signature is Optional[str] and None is a valid value (it produces an unnamed DataArray). slope.py has the same str vs Optional[str] docstring wording, so this matches the rest of the codebase. Not worth changing on its own.

What looks good

• The name= default of 'viewshed' is the same string the RTX path used to hardcode, and it follows the sibling-module convention (slope, curvature, hillshade all default to their own function name).
• name= reaches every backend helper (_viewshed_cpu, _viewshed_windowed, _viewshed_dask, viewshed_gpu, _viewshed_rt) and is set at DataArray construction. Verified live: numpy, cupy/RTX, dask+numpy, and dask+cupy all come back with name='viewshed' (or the custom name). Passing name= at construction does override the dask graph token here, so the dask paths no longer leak it.
• The float64 cast widens from float32, so values do not change. The numpy/cupy value differences are the pre-existing algorithm difference already noted in the docstring, not something this PR introduced.
• Tests cover all four backends with and without max_distance, plus a custom-name case. cupy and dask+cupy are gated on has_rtx().

Checklist

• Algorithm unchanged (metadata-only change)
• All four backends produce consistent name and dtype
• NaN handling unchanged
• Edge cases: existing suite still passes (37 tests)
• Dask chunk handling unchanged
• No premature materialization introduced
• Benchmark not needed (no perf-critical change)
• README feature matrix not applicable (no new function, no backend change)
• Docstring present for the new parameter