feat: number writes with store revisions

[skyoo2003]

Summary

Every write that changes the keyspace now gets a revision, so callers can tell which write produced the state they see. Data directories move to format 2. Format 1 directories are upgraded in place.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• Refactor
• CI/CD

Changes

• Store.Revision and Store.WriteRevision report the store's revision. A transaction is one revision however many keys it touches; a transaction that changes nothing does not advance it.
• Entry gains CreateRevision and ModRevision. Recreating a deleted or expired key counts as a new creation.
• The append log records revisions on set, del, and flush, and a new rev record heads every snapshot and log rewrite, so a store reopens at the revision it stopped at rather than one past whatever survived compaction. Records without revisions (format 1) load at revision 1.
• In a cluster, a write's revision is the Raft index of the entry that carried it, so every node agrees. ApplyReplicated takes that index and ignores the leader's speculative revisions; ReplaceWith restores the snapshot's revision, including zero for an empty snapshot. SetReplicator and the FSM now return the committed revision.
• internal/datadir: format version 2, which also reads format 1 and restamps it.
• Docs: clustering notes on in-place upgrades and rolling upgrades; compatibility page updated for the new API.

Changelog

• changie new fragment added — one sentence, two at most, 280 characters max
• Not user-visible, no fragment needed

Testing

• go test ./... passes
• New tests added (if applicable)

New tests in pkg/kvs/revision_test.go cover revision counting, restart persistence, legacy logs, replicated revisions, and snapshot restore (including an empty snapshot). internal/cluster and internal/datadir tests cover the Raft-index revision and the format 1 to 2 upgrade.

Related Issues

None