Fix cumulative_viewshed TypeError on cupy rasters (#3194)

[brendancol]

Fixes #3194

Problem

cumulative_viewshed and visibility_frequency raised
TypeError: Unsupported type <class 'numpy.ndarray'> on a cupy-backed raster.
viewshed() returns a cupy-backed DataArray for cupy input, but the accumulator
was a numpy array (count = np.zeros(...)), and cupy refuses a numpy+cupy add.
The cupy path had no test coverage, so this went unnoticed.

Fix

Pull each per-observer viewshed result to numpy with .get() before
accumulating into the numpy count. This matches how _extract_transect
already handles cupy data in the same module.

Test

Added a cupy regression test (gated on rtxpy) confirming both functions run
and return valid int32 counts. It does not assert cell-for-cell parity with
numpy: the GPU viewshed uses RTX ray tracing and the numpy backend uses the
GRASS sweep, a divergence already documented on viewshed.

Verified on an RTX A6000: both functions run, return int32 / float64 with
valid ranges, 26 visibility tests pass.

Found by the accuracy sweep (Cat 5: backend inconsistency).

[brendancol]

Closing in favor of #3205, which fixes the same cumulative_viewshed cupy crash but keeps the accumulator on-device, so cupy input yields a cupy-backed result (consistent with the cross-backend parity convention). This PR's .get() approach returns numpy output for cupy input, and the two test suites assert contradictory result types, so they can't both land. #3205 also covers visibility_frequency and metadata preservation.