[jayzhan211]

Which issue does this PR close?

• Part of [EPIC] Simplify memory-limited (spilling) execution #26711.

Stacked on #26987. Only the last commit, "refactor: one admission rule for every spill merge", is new here; I'll rebase once #26987 merges.

Rationale for this change

After #26987, the spill merge still has two options that only the aggregate sets: with_replay_headroom (#25383) and with_intermediate_merge_sizing (#25428). #26711 asks to remove options like these, and #26264 proposes a third, with_fan_in_headroom, so that a spilled sort also leaves memory for other operators.

This PR replaces all three with one rule that every caller gets: each run a merge reads must leave as much again as the merge takes beyond the workspace it already holds. This is the formula from #26264.

• The aggregate holds no workspace, so it gets its replay headroom as before.
• The sort holds sort_spill_reservation_bytes, so only what it takes beyond that is doubled.

What changes are included in this PR?

• The admission rule above, for sort and aggregate. With one batch read ahead, the two runs a pass needs are admitted against the pool alone, as in fix: write sorted spill runs in byte-bounded batches and drop the halving re-spill #26987. fix: leave memory for others when a spilled sort merges its runs #26264 exempts that minimum at any read-ahead.
• Intermediate merge sizing, which merges only as many runs as the final pass needs, now applies to every caller. It is estimated with the widest run and capped by max_spill_merge_fan_in, replacing the full-batch and uniform-run checks.
• Removed: with_replay_headroom and with_intermediate_merge_sizing on both merge builders, AggregateSpill's minimum batch-row tracking, and the row count returned by the intermediate spill.

What is the testing strategy for this PR?

• Merge unit tests that set the two options now run without them. The sizing tests check the same outputs and that sizing never rewrites more rows than merging every admitted run.
• Local runs (all pass): physical-plan lib tests, core memory_limit tests, extended fuzz tests, and the full sqllogictest suite.
• Not done yet:
  • sort_tpch under memory limits, which is needed to judge the cost below.
  • fix: leave memory for others when a spilled sort merges its runs #26264's concurrent-partition reproducer. fix: leave memory for others when a spilled sort merges its runs #26264 reports that its policy takes that query from 25/40 to 40/40 successes with sort_spill_reservation_bytes = 1 MB, and from 7/40 to 35/40 at 3 MB.

Are there any user-facing changes?

Query results don't change. A spilled sort's merge now leaves headroom beyond its workspace, as #26264 proposes. This applies even when nothing else is using memory, so it can add merge passes. Sorting 3M rows by an integer and a short string on one partition, with sort_spill_reservation_bytes = 4 MB:

Memory limit	Rows written to spill files, main	This PR
8 MB	8.95M	11.62M
24 MB	3.00M	4.80M

Intermediate merge sizing now applies to sorts as well as aggregates.