Add native physical dispatch for planned GFQL routes - #1448

Closes #1441.

Summary

• route planned SamePathExecutorWrapper and RowPipelineExecutorWrapper through a native chain-dispatch helper instead of _execute_compiled_query_compat_non_union
• keep the compatibility wrapper for unplanned compiled queries and CALL-backed planner gaps
• add runtime cutover guards for natural same-path and row-pipeline Cypher queries
• harden RAPIDS 25.02 GFQL crash lanes found by full GFQL DGX validation:
  • mixed pandas/cuDF row projection now bridges through Arrow instead of cuDF to_pandas()
  • radial layout GPU trig on RAPIDS 25.02 uses a host trig bridge instead of cuDF to_cupy()

Validation

• python3 -m pytest -q graphistry/tests/compute/gfql/test_runtime_physical_cutover.py graphistry/tests/compute/gfql/test_physical_planner.py -> 13 passed
• python3 -m pytest -q graphistry/tests/compute/gfql/layouts/test_call_layout_radial.py graphistry/tests/compute/gfql/test_runtime_physical_cutover.py graphistry/tests/compute/gfql/test_physical_planner.py -> 22 passed
• python3 -m pytest -q graphistry/tests/compute/gfql/cypher/test_lowering.py -k "not cudf" -> 930 passed, 12 skipped, 108 deselected
• python3 -m pytest -q graphistry/tests/compute/gfql/test_row_pipeline_ops.py -k "not cudf" -> 182 passed, 16 deselected
• ./bin/ruff.sh graphistry/layout/ring/util.py graphistry/compute/gfql/row/pipeline.py graphistry/compute/gfql_unified.py graphistry/tests/compute/gfql/test_runtime_physical_cutover.py -> passed
• DGX RAPIDS 25.02 full GFQL directory:
  RAPIDS_VERSION=25.02 PROFILE=gfql WITH_IMAGE_BUILD=0 IMAGE_TAG=graphistry/test-rapids-official:25.02-cuda12-gfql TEST_FILES='graphistry/tests/compute/gfql' docker/test-rapids-official-local.sh
  -> 2249 passed, 22 skipped, 15 xfailed
• DGX RAPIDS 26.02 full GFQL directory:
  RAPIDS_VERSION=26.02 PROFILE=gfql WITH_IMAGE_BUILD=0 IMAGE_TAG=graphistry/test-rapids-official:26.02-cuda12-gfql TEST_FILES='graphistry/tests/compute/gfql' docker/test-rapids-official-local.sh
  -> 2249 passed, 22 skipped, 15 xfailed

Notes

• local full cuDF suites still cannot start on this host because cuDF imports but there is no CUDA device (cudaErrorNoDevice); DGX covers those lanes.