Return all-NaN from HPA* fallback when the goal cannot be reached

[brendancol]

Closes #3631

• _hpa_star_search now returns a fresh all-NaN surface when a segment cannot be refined at the largest retry radius, instead of the partially written path image. An unreachable goal no longer produces a finite cost trail that dead-ends mid-grid, so np.isfinite(result).any() is a reliable reachability check in the HPA* regime, same as in the plain and bounded searches.

Backend coverage: HPA* runs on the numpy path (and the cupy CPU-fallback path); the dask backends use the sparse A* and are unaffected.

Test plan:

• New test: wall-split 200x200 grid where the coarse route crosses the wall but refinement cannot; asserts the output has no finite pixels (previously 89)
• pytest xrspatial/tests/test_pathfinding.py (52 passed on a CUDA host)

[brendancol]

PR Review: Return all-NaN from HPA* fallback when the goal cannot be reached

Blockers (must fix before merge)

• none

Suggestions (should fix, not blocking)

• none

Nits (optional improvements)

• xrspatial/pathfinding.py:520 a refinement failure now discards segments that were already connected, so a caller gets no hint of how far the search got. That matches the bounded-search behavior codified in test_search_radius_too_small (all-NaN when the search bounds exclude the only route), and the partial trail was worse: it made np.isfinite(result).any() a false reachability signal. If partial progress ever matters, a warning would be the way to surface it, not finite pixels.
• The test drives _hpa_star_search directly rather than through a_star_search with a mocked _available_memory_bytes and a >1000-pixel manhattan distance. Direct invocation keeps the test fast and deterministic; the public-API route to this code path is already covered by test_auto_radius_selection.

What looks good

• One-line semantic fix with the reasoning in a comment at the return site.
• The new test constructs the exact failure geometry (thin NaN wall that coarse blocks see as passable) and asserts zero finite pixels, which fails on the old code with 89.
• No behavior change for reachable goals: the escalation loop and stitching are untouched, and the existing test_hpa_star_correctness still passes.

Checklist

• Algorithm matches reference/paper (no-path contract now uniform across search paths)
• All implemented backends produce consistent results (HPA* runs on the numpy path and cupy CPU-fallback; dask uses sparse A*, unaffected)
• NaN handling is correct
• Edge cases are covered by tests
• Dask chunk boundaries handled correctly (not touched)
• No premature materialization or unnecessary copies
• Benchmark exists or is not needed (failure-path fix)
• README feature matrix updated (n/a)
• Docstrings present and accurate (internal function; public docstring already describes HPA* only at a high level)