Balance spine batch events; document the stamp invariant

[frankmcsherry]

Addresses the differential items from Materialize's report on the master-next port that we chose to take.

Changes

• The spine merges two present spans without work when at least one carries no batch. Such a merge is born complete, and complete_at logged no completion for it. A consumer adding one batch per BatchEvent and removing one per completed MergeEvent and per DropEvent (as diagnostics does) counted those spans forever. insert_at now logs the completion, which restores the events of 0.25.1.
• The logging types state that contract. Batches without updates are counted, and report a length of zero. A consumer that counts only batches with updates (as Materialize now does) also balances.
• Collection and Arranged::stream state the unchecked assumption that each update's time is greater or equal to some time in its message's stamp, so a message with an empty stamp carries no updates.
• OrdValBuilder::with_capacity and OrdKeyBuilder::with_capacity are public, so wrapping builders can pre-size.
• Corrects two comments that described one capability per message, and one that named a caller drop_batches no longer has.

Tests

tests/logging.rs arranges 100 rounds, most without updates, drops the trace, and checks that both consumers return to zero. Without the spine change the first consumer ends at 79.

🤖 Generated with Claude Code