reproject test-coverage: close bounds_policy + Inf backend gaps

[brendancol]

Summary

Pass 2 of the test-coverage sweep on xrspatial.reproject (the 2026-05-10
first pass closed the bulk of the gaps; this catches the two backend-coverage
holes the new bounds_policy parameter and the existing Inf test left open).

10 new tests in xrspatial/tests/test_reproject_coverage_2026_05_27.py,
all passing on a CUDA host. Source untouched.

Closes two Cat 1 MEDIUM backend-coverage gaps:

1. bounds_policy (added in reproject: output-grid bounds use silent heuristics that can crop real data near singularities #2187) had numpy + dask+numpy coverage but no
   cupy / dask+cupy tests. A regression that dropped the kwarg from either
   GPU dispatcher would ship undetected. TestBoundsPolicyCupy and
   TestBoundsPolicyDaskCupy pin raw, clamp, and the bogus-token
   rejection on both GPU backends, plus a clamp-grid parity check against
   the numpy path.

2. test_reproject_handles_inf_input only covered eager numpy. The dask,
   cupy, and dask+cupy chunk workers each ship their own bilinear / cubic
   resampler so a regression raising on +/-Inf in any one backend would
   not surface from the existing test. Four new tests close the matrix:
   dask+numpy, cupy, and dask+cupy with scattered +/-Inf cells,
   plus cupy with an all-Inf raster pinning that no spurious finite
   cells appear in the output.

State CSV row updated to last_inspected=2026-05-27.

Test plan

• All 10 new tests pass locally on a CUDA host (pytest xrspatial/tests/test_reproject_coverage_2026_05_27.py -v)
• Mutation check: bypassing _validate_bounds_policy in reproject() flips the two bogus-token rejection tests red
• Full reproject test suite still passes
• CI green on the project's GPU runner