perf: send one AppendEntries RPC per matching-point probe

[zhixinwen]

Matching-point discovery drains every storage-limited chunk of a probe range before choosing the next midpoint. For example, a 500-entry probe split into ten storage reads can require ten RPCs even though the first response provides enough information to continue discovery.

This change sends one entry-carrying AppendEntries request per probe. Storage may return a prefix of the candidate range; once the follower acknowledges an entry beyond prev, the engine finishes the probe and recomputes the next midpoint. The existing max_payload_entries cap still applies.

The engine reuses Inflight::Logs: both probes and the leader's local progress update complete on the first acknowledgement beyond prev (the local update acknowledges the entire range immediately). Payload::Probe gives the replication task the probe-specific sending and retry rules, separate from commit-only appends. Completion is handled centrally by Payload::update_matching(), so next_request needs no probe-specific branch. Pipeline replication is unchanged, and all changed types are internal.

The engine and replication task share LogIdRange::probe_completed_by(). PartialSuccess(prev) verifies prev but acknowledges no additional entry, so the same probe is retried. A session with no response also leaves it pending. This decision uses the current session's maximum acknowledgement, since remote_matched can contain progress from earlier sessions. Keeping the maximum prevents a later, lower response from undoing the session's completion decision.

Tests cover:

• A storage reader returning entry 53 for (52, 60]: the request stream ends without reading or sending the suffix. An initial empty storage read leaves the probe pending.
• A prefix acknowledgement completing the inflight probe and causing the engine to choose a new midpoint.
• A fresh-learner probe (prev = None) remaining pending without an acknowledgement and completing when the first entry is acknowledged.
• Zero-entry partial responses: both followers must issue a second entry-carrying request before the test lifts the quota, then catch up. Synchronization uses per-follower retry notifications.

A transport closing its response stream without answering is not directly covered by a network mock.

Validation: formatting and all three workspace Clippy configurations pass. make verify encounters a local C++ standard-header error while building the unchanged RocksDB example. make -o fmt_check -o clippy verify passes documentation checks, library tests, integration tests, and the remaining macro checks.

---

This change is 

[zhixinwen]

I am reopening this PR. As commented in #2073 (comment), changing size to 8 can cause regression to other users and it makes the binary search within small gap slower. Even with #2076, we would still need binary search so this optimization would still help.

[drmingdrmer]

Thank you, @zhixinwen, for identifying this issue, implementing the optimization, and asking us to help refine it.

The core idea remains unchanged: once a probe sends an entry-carrying request, its response contains enough information to choose the next midpoint. Sending the rest of the candidate range only delays matching-point discovery.

We made a follow-up refactor on foo to make this behavior easier to reason about:

• Removed Payload::LogIdRange and renamed Inflight::Logs to Inflight::Probe.
• Kept LogIdRange as a data type. Probe-specific completion logic now belongs to Payload and Inflight.
• Made Payload::update_sent() and update_matching() return the remaining payload. A sent range is never returned for another send.
• Defined probe completion as acked >= prev, including an empty acknowledgement for a fresh learner.
• Changed SessionOutcome to contain only the acknowledgement and remaining payload. Option<Option<LogId>> distinguishes no response from an acknowledgement of None.
• Simplified response handling to return the greatest acknowledgement and the reason for stopping.
• Unified handling of Some(log_id) and None acknowledgements, and removed the duplicate remote_matched update.
• Updated the unit and integration tests for limited storage reads, empty probes, partial success, and choosing the next midpoint.

This preserves the optimization while making the probe lifecycle and retry rules more explicit. make verify passes.

[drmingdrmer]

@drmingdrmer reviewed 23 files and all commit messages.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on zhixinwen).

[drmingdrmer]

@drmingdrmer reviewed 5 files and all commit messages.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on zhixinwen).

[drmingdrmer]

@drmingdrmer reviewed 5 files and all commit messages.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on zhixinwen).