Fold pipelineable downstream ops into merge pipelines after GROUP BY and TOP_N

[Jedi18]

Summary

After HASH_GROUP_BY / TOP_N splitting, Sirius previously created a small merge pipeline (PARTITION → MERGE_*) and then a separate downstream pipeline for streaming ops (typically RESULT_COLLECTOR). That extra pipeline boundary is unnecessary when downstream is a dead end.

This PR folds pipelineable downstream operators into the merge pipeline when safe, so merge and collector run in a single gpu_pipeline_task:

Before: PARTITION → MERGE_GROUP_BY | RESULT_COLLECTOR
After: PARTITION → MERGE_GROUP_BY → RESULT_COLLECTOR

Fusion is applied from split_group_aggregate_sink and split_top_n_sink via attach_merge_and_fuse_downstream, with three guards in try_fuse_downstream_pipelineable:

• Guard 1: block when a direct downstream pipeline ends in a structural sink (ORDER BY, HASH_JOIN, nested GROUP BY, etc.) that needs its own split_* path
• Guard 2: block when a later structural pipeline reads from the candidate downstream’s output (e.g. GROUP BY → PROJECTION → HASH_JOIN) — fusion would remove a pipeline boundary that join wiring still depends on
• Guard 3: only absorb pipelines whose operators are plain intermediates (!is_source() && !is_sink())

UNGROUPED_AGGREGATE is explicitly excluded: the same physical op is both the partial-aggregate sink and the merge pipeline source. Folding downstream broke partial→merge repository wiring and produced wrong results (e.g. avg off by orders of magnitude).

Adds absorbed_pipelines_ so absorbed downstream pipelines are skipped in split_pipelines() and not double-scheduled.

Also includes regression tests in test_pipeline_merge_fusion.cpp

[bwyogatama]

The query planner is undergoing major refactoring and split_group_aggregate_sink and split_top_n_sink will be deleted after this PR #790.
So I would work on this after that PR is in.

[Jedi18]

Thanks for letting me know @bwyogatama, I'll put this PR on hold till then I guess