Route the HPA* coarse grid without value barriers

[brendancol]

Closes #3647

• _hpa_star_search now routes the coarsened grid with an empty barrier list. Coarse cell values are block means of the non-barrier fine cells, so testing them against the exact barrier values could falsely block a passable block whose values average to a barrier value (alternating -1/+1 data with barriers=[0] makes every block mean exactly 0.0, blocking the whole coarse grid and returning all-NaN "no path" on a fully passable surface).
• _coarsen_surface already marks fully-barrier blocks as NaN, and NaN cells stay impassable in the coarse search, so no blocking information is lost. Refinement still runs on the fine grid with the caller's barriers.

Backends: HPA* runs on the numpy and cupy (CPU-fallback) paths only; the dask sparse A* never coarsens.

Test plan:

• test_hpa_star_barrier_valued_block_means (fails on main: 0 finite pixels)
• Re-ran the issue's end-to-end repro (600x600, mocked RAM, auto-radius HPA*): goal reached after the fix
• pytest xrspatial/tests/test_pathfinding.py — 58 passed, including the existing HPA* correctness and all-NaN no-path tests
• flake8 clean on changed lines

Also includes this sweep's state-CSV update.

https://claude.ai/code/session_0155N4QGamQVxgpAAPbpQNq4

[brendancol]

Domain-aware review of the HPA* coarse-barrier fix.

Blockers

None.

Suggestions

• The empty array reuses barriers.dtype, which keeps the numba signature of _a_star_search stable across the coarse and fine calls when callers pass integer barrier lists. Good; a bare np.array([]) would have forced a second float64 specialization but also worked. No change needed.

Nits

• The fix only touches the coarse route. Refinement (_bounded_a_star_sub) still runs on the fine grid with the caller's barriers, so true barriers keep constraining the actual path; the change removes only the mean-aliasing false blocks. The pre-existing semantics that a block with any single crossable cell is coarse-passable are unchanged, and the radius-escalation loop plus the #3638 all-NaN contract still cover refinement dead-ends.

Verified locally: the new test returns 0 finite pixels on main and a connected start-to-goal path with the fix; the issue's end-to-end repro (600x600, mocked RAM so auto-radius HPA* engages, barriers=[0]) reaches the goal after the fix; test_hpa_star_correctness and test_hpa_star_unreachable_goal_all_nan still pass (58 total).