Nearest-center point mapping and out-of-bounds rejection in pathfinding _get_pixel_id

[brendancol]

Closes #3629

• _get_pixel_id now computes a signed offset from coords[0] (handles ascending and descending coords) instead of abs(), so a point outside the raster gets an out-of-range index and a_star_search / multi_stop_search raise their existing "outside the surface" errors on every side instead of silently pathing from a mirrored interior pixel.
• The index is rounded to the nearest cell center instead of truncated. A point in the upper half of a cell no longer shifts down one pixel, and a user-typed 0.3 on a 0.1-resolution grid (pixel-3 center 0.30000000000000004) resolves to pixel 3 rather than pixel 2.
• Also carries the accuracy-sweep state CSV row for the pathfinding module (findings a_star_search maps out-of-bounds points to interior pixels and truncates instead of rounding in _get_pixel_id #3629, multi_stop_search crashes on dask+cupy input and returns numpy-backed output for cupy input #3630, HPA* fallback in a_star_search returns a partial cost trail when no path exists #3631).

Backend coverage: _get_pixel_id runs on the host for every backend, so the change applies uniformly to numpy, cupy, dask+numpy, and dask+cupy inputs.

Test plan:

• New tests: nearest-center mapping (descending and ascending coords, float-noise case), out-of-bounds raises on all four sides for start and goal, out-of-bounds waypoint raises in multi_stop_search
• pytest xrspatial/tests/test_pathfinding.py (55 passed, includes cupy and dask+cupy on a CUDA host)
• pytest xrspatial/tests/test_cost_distance.py (89 passed)

[brendancol]

PR Review: Nearest-center point mapping and out-of-bounds rejection in pathfinding _get_pixel_id

Blockers (must fix before merge)

• none

Suggestions (should fix, not blocking)

• xrspatial/pathfinding.py a_star_search docstring (start/goal params): the point-to-pixel convention is now nearest-cell-center with out-of-bounds rejection on every side. Worth a sentence in the docstring, since this PR also tightens behavior for points more than half a cell past the far edge (previously tolerated up to a full cell via truncation, now a ValueError).

Nits (optional improvements)

• xrspatial/pathfinding.py:46 sign detection comes from the coords while the cell size still comes from get_dataarray_resolution (which prefers attrs['res']). If attrs['res'] disagrees with the actual coord spacing the mapping is off, but that trade-off predates this PR and is consistent with the rest of the module.
• Python round() uses half-to-even, so a point exactly halfway between two centers snaps to the even index. Deterministic and either neighbor is defensible; noting it so nobody reads it as a bug later.

What looks good

• The signed-step change makes the existing _is_inside validation actually fire on the coords[0] side; the fix reuses the module's own error paths instead of adding new checks.
• Tests cover descending and ascending coords, the float-noise case at an exact center (0.3 on a 0.1-res grid), all four out-of-bounds sides for both start and goal, and the multi_stop_search waypoint check.
• Existing suite still passes (55 tests including cupy and dask+cupy on a CUDA host), and the fixture points in older tests sit exactly on cell centers, so no codified behavior changed.

Checklist

• Algorithm matches reference/paper (nearest-center mapping, verified by hand cases in tests)
• All implemented backends produce consistent results (mapping runs host-side for all four)
• NaN handling is correct (unchanged by this PR)
• 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 (scalar point mapping, not benchmark-relevant)
• README feature matrix updated (n/a, no new function)
• Docstrings present and accurate (see suggestion about documenting the convention)

[brendancol]

PR Review follow-up

The one suggestion from the first pass is addressed in 481b69e: the start/goal parameter docs now state the nearest-cell-center mapping and that out-of-bounds points raise ValueError. Both nits were informational (pre-existing attrs['res'] precedence, half-to-even rounding at exact midpoints) and need no code change. Tests re-run after the docstring commit: 55 passed. No open findings.