feat: event-driven projector, connection pooling, async writes, handler registration, WriteBatch API

[johnny-emp]

Summary

• Event-driven projector: FileEventLog fires tokio::sync::Notify after each successful append. Projector::run_continuous() and EventProcessor::run() now await the notification instead of blind-polling, giving near-zero-latency projection with zero CPU waste when idle. Falls back to poll_interval timeout as a safety net.
• Projection read pool enabled by default: ProjectionConfig now creates a 4-connection read pool out of the box. Added projection_write_conn() / write_conn() for explicit write access; projection_connection() documented as deprecated.
• Native async write API: LmdbCanonicalStore::submit_write() centralizes spawn_blocking + transaction lifecycle into a single closure-based method. Convenience wrappers async_put_state() and async_del_state() for common single-op writes. Exposed via AzothDb::submit_write().
• Handler registration: EventProcessorBuilder::build_with_projection() opens a dedicated read-write SQLite connection to the projection DB automatically, removing the need for callers to pass one externally.
• WriteBatch API: New WriteBatch struct for explicit multi-operation atomic commits with put(), delete(), append_event(), commit(), and commit_async(). Accessible via db.write_batch().

Test plan

• cargo check passes across the full workspace
• cargo test --workspace passes (all existing tests green, no regressions)
• No linter errors on any modified files
• Confirm read pool default doesn't break existing consumers that rely on disabled pool