Optimize filterless group contribution caps

[ila]

Thanks for reviewing.

Summary

• Replace the filterless duplicate-aware DENSE_RANK group cap with ROW_NUMBER after the unique (PU, group) pre-aggregation.
• Order by a stable hash(PU, group), matching the bounded-selection strategy already used by ordinary standard DP and SASS aggregates.
• Verify that DuckDB rewrites the cap to a bounded per-PU top-k aggregate.

Why this is safe

The lower aggregate emits exactly one row per logical (PU, group), so duplicate-aware ranking is unnecessary. ROW_NUMBER <= C_u still retains at most C_u groups per PU. A hash collision can change which row wins, but cannot admit more than the cap. Distinct-value paths retain DENSE_RANK, where duplicate logical items must share a rank.

Validation

• clean Ninja build from current main
• build/release/test/unittest test/sql/dp_filterless.test (140 assertions)
• C_u=1 and C_u=2 correctness coverage
• C_u=3 with 100 PUs and 10 candidate groups: exactly 300 contributions retained
• physical-plan regex requires UNNEST over arg_min_nulls_last for the C_u>1 top-k shape
• real 1,000,000-row grouped C_u=3 query also retained exactly 300,000 contributions