fix(gfql): the fused grouped-aggregate lane keeps the group_by plan on the GPU target (#2064)

[lmeyerov]

Closes the fused-lane half of #2064. Independent of the release stack (base master).

What it was. Not a dtype and not a GPU capability gap in our checks: the fused single-hop grouped-aggregate lane's low-cardinality count(*) alternative lowers to col(key).value_counts() + unnest, and cudf-polars 26.02 has no unnest map function (NotImplementedError: Unhandled map function unnest from raise_on_fail=True). _run_fast_path_on_requested_target converted that into "declined; plan not GPU-executable", so on engine='polars-gpu' every fused count(*) shape fell to the generic route; the values were right (the generic route answered), and the 71 + 8 + 2 + 2 fused pins failed on their served-by assertion. Found by dumping the failing frame's plan at the lane's collect on dgx.

Fix. _low_cardinality_pure_count_plan declines on the GPU target, so the lane keeps the group_by(...).agg(pl.len()) formulation there, which cudf-polars executes (verified directly on the engine). CPU keeps the value_counts plan unchanged.

Pin. test_low_cardinality_count_plan_declines_on_the_gpu_target: under target_mode(GPU) the alternative returns None; under CPU it still yields the UNNEST plan. Runs without a GPU.

Not in scope. The test_strictness_levels.py polars-gpu cases fail on the CPU polars engine too (select rows op not natively supported on polars); and the absent-property placeholder in gfql_unified.py:340 is an untyped pl.lit(None), which cudf-polars rejects (Invalid type_id) once that path is reachable; both stay on #2064.

Receipts: local fused file 252 pass; guards + mypy green; GPU receipt below.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QztW7jYsDd66e8rb8pJNQA

[lmeyerov]

GPU receipt (dgx-spark, graphistry/test-rapids-official:26.02-gfql-polars, cudf 26.02.01 / polars 1.35.2, safe_run, TEST_CUDF=1) at a4b6b4b on test_grouped_aggregate_fused_polars.py + test_strictness_levels.py: 607 passed / 5 failed, versus 520 passed / 91 failed on master 86de0f5 for the same two files. Every fused grouped-aggregate pin now passes on polars-gpu; the 5 left are the strictness polars-gpu cases (MATCH (n) RETURN n.nope_col), which raise the polars engine's select rows-op decline on CPU polars too and stay on #2064.

[lmeyerov]

CI receipt at a4b6b4b: 65 check-runs success, 4 path-skipped (tck-gfql, test-polars ×6, test-gfql-core, gfql-benchmarks and the guards green). GPU receipt above (607 / 5 vs master 520 / 91 on the two files).

[lmeyerov]

CI receipt at e1eb737 (union-rebased onto master after #2055 landed; own delta unchanged): 65 check-runs success, 4 path-skipped. GPU receipt above stands (the change is in the lane's plan choice only).

[lmeyerov]

Folded into #2062 (owner asked for a shallower stack); the commits, receipts and CHANGELOG entries live there unchanged.