compute: park future updates in a merge batcher wrapper - #39928

`TemporalBatcher` wraps any differential `Batcher` and keeps updates at
least θ past the last sealed frontier in the existing bucket chain of
time-major buckets, so the inner batcher does not re-merge them on every
seal. Seals peel the chain at the frontier's inner-minimum key and feed
the peeled updates through a chunker into the inner batcher, which
decides readiness. The first seal rebuilds the inner batcher, because
the snapshot and its future tail arrive before the batcher learns the
frontier.

Buckets are keyed by a new `BucketTime` trait, the identity for totally
ordered times, and track a lower bound of their times for the wrapper's
frontier. The wrapper reads output chunks through `TimedUpdates`,
implemented for the three output types of the arrange batcher arms.

Adds model tests against the plain batcher (hand-written, hydration
shaped, spilled, and proptest) and a criterion bench.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…ing an empty chain

Co-Authored-By: Claude Opus 5.5 <[email protected]>

At the ArrangeBy ok arrangement, wrap all three batcher arms in
`TemporalBatcher` when the timestamp is `mz_repr::Timestamp`, and stop
applying the temporal bucketing operator in `ensure_collections`.
Iterative timestamps keep the plain batchers. `apply_worker_config`
installs θ from `compute_temporal_bucketing_summary`.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Implement `BucketTime` for `Product` by its outer coordinate, and use
the batcher wrapper for iterative timestamps at the ArrangeBy ok
arrangement too. A seal parks updates only relative to the least key
whose inner-minimal time is in the frontier, so updates that might be
ready always sit in the inner batcher. Adds model tests over
`Product<u64, u64>` times.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Co-Authored-By: Claude Opus 5.5 <[email protected]>

The temporal batcher marked itself rebuilt at its first seal even when
that seal's upper had no inner-minimum element, so a later seal with a
far bound would leave the hydration tail in the inner batcher. Rebuild
at the first seal that has a far bound instead. With advancing uppers
the case cannot arise, since an inner-minimum element only follows
another, so the change is defensive. Rename the stats to match.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…ites

Replace the temporal bucketing operator with batchers that park future
updates at Reduce input arrangements and the monotonic consolidation,
TopK arrangements and consolidations, linear join stage arrangements,
the iterative index export, and `columnar_consolidate`, which covers
consolidating Unions and `LetRec` bindings. `RenderTimestamp` now
requires `BucketTime` keyed by `mz_repr::Timestamp`, so the wrapped
batchers need no per-site dispatch.

The accumulable Reduce plan keeps the operator: its arrangement carries
`(Vec<Accum>, Diff)` diffs, which the bucket storage cannot consolidate
because the diff does not add a columnar ref. The lowering's
`TemporalBucketing` strategies stay plumbed but mostly go unused.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Match `compute_temporal_bucketing_summary` and timely's `Summary`
naming in fields, setters, the process-global, comments, tests, and the
bench.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

The batcher inferred hydration from its first seal and rebuilt its inner
batcher there, moving the snapshot's future tail into the bucket chain.
The dataflow builder now publishes the dataflow `as_of` in a scoped
thread-local that `TemporalBatcher::new` reads. With it, the batcher
classifies pushes before its first seal against `as_of` plus the
summary, so the tail parks on arrival and the snapshot stays in the
inner batcher, and it joins each seal's edge with `as_of` as the
bucketing operator does. The chain is shaped around `as_of` up front,
so the tail does not land in one bucket spanning the domain. The
rebuild and its stats go away. Without an `as_of` nothing parks before
the first seal.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Every site that bucketed a columnar edge now parks future updates in its
batcher, so `MaybeBucketByTime::maybe_apply_temporal_bucketing` has no
callers. The `Vec` variant stays for the accumulable reduce path.

Co-Authored-By: Claude Opus 5.5 <[email protected]>