[AURON #2507] Optimize partial WindowGroupLimit execution

[lyne7-sc]

Which issue does this PR close?

Closes #2507

Rationale for this change

Partial WindowGroupLimitExec currently materializes and filters a ranking column that is not included in its output. It also always uses Arrow filtering even when the selected rows form a contiguous range.

What changes are included in this PR?

• Add a specialized partial group-limit processor for row_number, rank, and dense_rank.
• Track partition, ordering, and ranking state across input batches.
• Collect adjacent selected rows as ranges instead of building the selection bitmap row by row.
• Use zero-copy RecordBatch::slice for a single contiguous range.
• Build the Boolean selection in blocks for multiple ranges.
• Add correctness tests and operator-level benchmarks for contiguous and multi-range selections.

Are there any user-facing changes?

No. This is an internal performance optimization.

How was this patch tested?

Passed unit tests.

Benchmarks

The benchmark measures WindowGroupLimit execution over 100,000 input rows. Input sorting is excluded from the measurement.

Benchmark	Master	This PR	Improvement
window_group_limit_row_number	17.558 ms	13.538 ms	22.9%
window_group_limit_rank	32.092 ms	27.360 ms	14.7%
window_group_limit_dense_rank	32.683 ms	28.133 ms	13.9%
window_group_limit_row_number_single_partition	17.612 ms	13.175 ms	25.2%
window_group_limit_row_number_all_selected	17.436 ms	14.519 ms	16.7%

Was this patch authored or co-authored using generative AI tooling?

• Yes
• No

If yes, include: Generated-by: GPT-5

ASF guidance: https://www.apache.org/legal/generative-tooling.html

[slfan1989]

Thanks for this optimization.

The specialized processor correctly maintains partition, ordering, and ranking state across batches, and the range-based filtering is well scoped.

The tests cover row_number, rank, and dense_rank, including peer groups spanning batch boundaries, and the CI checks are green.

Two minor documentation nits:

• The benchmark table appears to be out of sync with the checked-in benchmark names. It lists window_group_limit_row_number_single_partition and window_group_limit_row_number_all_selected, while the current benchmark file defines the *_slice cases instead.

• Please include the actual tool name in the AI disclosure, for example Generated-by: OpenAI Codex (GPT-5) if applicable.

No blocking code concerns otherwise. LGTM!