Move reduce timestamps and directly evaluate single dominating seeds

[frankmcsherry]

A key with one seed that dominates every selected input/output timestamp has exactly one possible evaluation. Reconcile its input/output ranges directly, or retain the seed if it has not crossed the frontier, while leaving the general partial-order walk for all other keys. This avoids building and replaying histories for the single-evaluation case.

Also move consumed timestamps through two ownership boundaries: selected rows into Corgi's presentation merge, and bridge rows into histories. Neither consumed source needs its old timestamp again. The production diff is +60/-35 (+25 net) across int_proxy reduce and Corgi presentation.

Three focused tests (63 lines) exercise four conditions: a seed beyond upper must remain pending; that seed subsequently becomes due with no novel input; input and compacted prior-output records exactly at the seed must participate; and an incomparable record requires the general walk, including a correction at the joined time. Existing partial-order/window-size, cancellation, collision and backend tests remain in place.

Validation: cargo test --locked --offline -p differential-dataflow -p interactive passed (196 regular tests and 54 doc tests; 8 regular and 6 doc tests remain ignored). The incomparable-record test uses a two-element compaction frontier. Validation was rerun after rebasing onto merged #881, including the grouped-pending tests and all three direct-evaluation tests.

Original isolated SCC measurement before the rebase, against #880, with neither grouped pending nor the history replacement on either side. These are not measurements of the incremental gain above #881:

	#880 baseline	Ownership moves + direct evaluation	Reduction
Startup	5.696 s	4.978 s	12.6%
Ten updates	7.039 s	6.458 s	8.3%
Peak RSS	1464 MiB	1454 MiB	0.7%

Apple M4, one worker, release/LTO, 500k nodes/1M edges, ten rounds of 1,000 replacements, three alternating process pairs, same harness/compiler/Timely 4eb3ee09681c209ebcaeced73548a9fbf8e1c71e. Output checksums match and there were no pageout/swap counter increases. These are measured combined effects of this PR, not sums of the individual experiments. The baseline executable was built from #880's reviewed head; its repository contents match merged dc3733d.

Based on master-next at 680500b (#881 merged). Direct evaluation reads due times through the grouped pending storage now in the base. The shared-history representation change in #882 is excluded.