Guard single-point input in kriging() variogram

[brendancol]

Closes #2917

kriging() crashed with an opaque numpy ValueError ("zero-size array to reduction operation maximum which has no identity") whenever the input reduced to a single point: either one point passed directly, or all but one point dropped by NaN filtering.

• _experimental_variogram now returns empty arrays when there are no distance pairs, before calling dists.max(). This routes the single-point case to the existing default-variogram fallback that already handles coincident points, so the call produces a constant-valued raster instead of raising.

Backend coverage: numpy (the variogram and matrix build run on host arrays for all four backends, so the fix applies regardless of the template's backend).

Test plan:

• test_single_point — one input point yields a finite constant prediction
• test_single_valid_point_after_nan_drop — NaN filtering down to one point reaches the same fallback
• existing kriging suite still passes (18 passed)

[brendancol]

PR Review: Guard single-point input in kriging() variogram

Blockers

None.

Suggestions

None.

Nits

None.

What looks good

• The guard is placed correctly: dists.size == 0 is checked before dists.max(), which is exactly where the empty-reduction crash originated (_kriging.py).
• It reuses the existing default-variogram fallback rather than adding a parallel code path. The dists.size == 0 (n<2) and max_dist <= 0 (coincident points) cases stay distinct, and both are needed.
• Because the variogram and matrix build run on host numpy arrays ahead of backend dispatch, the fix covers all four backends without touching the dispatch layer.
• Two tests cover both routes into the single-point state: one point passed directly, and one point left after NaN filtering. Both assert a finite constant prediction and the expected warning.

Verification

• Reproduced the original ValueError on main, confirmed it is gone after the change.
• Single point returns a constant raster equal to the point value (42.0 and 7.0 in the two tests); ordinary-kriging weight collapses to 1.0 with no spatial structure, which is the sensible degenerate result.
• Full kriging suite: 18 passed.

Checklist

• Algorithm correct for the degenerate case
• All backends covered (fix is upstream of dispatch)
• NaN handling correct (NaN-drop-to-one routes to fallback)
• Edge cases covered by tests
• No premature materialization
• Benchmark not needed (bug fix)
• README not applicable (no API change)
• Docstrings unchanged; behavior already documented via MemoryError/warning paths