Fix disaggregate(limiting_variable) silently losing a zone's value (#3403)

[brendancol]

Fixes #3403.

disaggregate(method='limiting_variable') dropped a zone's whole value when the zone had no pixel that could hold it. If every pixel in a zone fell in a cap-0 class (all zero-weight under the default caps), the leftover loop found no overflow pixels and threw the remaining mass away, leaving those pixels at their initialised 0. The value disappeared with no error, even though the docstring promises zone totals are preserved.

The weighted method already handles this by returning NaN. This change makes limiting_variable do the same: when no pixel can absorb the leftover, the zone's result is NaN instead of silent zeros.

zones = xr.DataArray(np.array([[1, 1], [1, 1]], dtype=np.float64), dims=['y', 'x'])
weight = xr.DataArray(np.zeros((2, 2)), dims=['y', 'x'])

# before: [0. 0. 0. 0.], nansum 0.0  (500 lost)
# after:  [nan nan nan nan]          (matches weighted)
disaggregate(zones, {1: 500.0}, weight, method='limiting_variable')

Added test_uninhabitable_zone_is_nan_not_zero, which fails on the old code and passes now. The full dasymetric suite stays green (62 tests, including the cupy and dask+cupy cross-backend checks on this GPU host). The cupy and dask backends route through the numpy core, so the fix reaches all four backends.

Found by the accuracy sweep on the dasymetric module.