fix(raft): make Txn sub-operations atomic via shared WriteBatch

[slow2342]

Summary

Txn sub-operations now share a single WriteBatch and commit atomically, matching etcd semantics. Previously each sub-operation committed independently, causing partial application on mid-sequence failures.

Changes

• src/raft/state_machine.rs — Core refactor of the apply path:

  • Added TxnSnapshot for rollback of in-memory state (key_indexes, key_metas, key_leases, lease_manager) on batch write failure
  • Extracted batch_put/batch_delete helpers that append to a shared WriteBatch without calling db.write()
  • Added key_leases in-memory map to track lease associations accurately within multi-op Txns
  • Added value_cache to fix prev_kv inconsistency when the same key is written multiple times in one Txn
  • Moved save_global_revision call out of sub-ops to caller level (called once per operation)
  • Added 3 new tests: test_txn_multi_put_atomic, test_txn_put_delete_atomic, test_txn_watch_events_emitted

• src/raft/node.rs — Persist and restore applied index:

  • Load applied_index from raft_state CF on startup
  • Skip entries with index <= applied_index to prevent re-apply on restart
  • Persist applied_index after applying each batch of committed entries

• src/raft/raftrs_store.rs — Applied index persistence:

  • Added save_applied_index() and load_applied_index() methods for the raft_state CF

Testing

• All 58 existing tests pass
• 3 new tests verify Txn atomicity (multi-put, put-delete, watch events)
• cargo clippy -- -D warnings passes
• cargo fmt -- --check passes

Related issues

Refs #108