Refactor WAL appends to async background writer with batching and backpressure

[bravo1goingdark]

Motivation

• Reduce contention on the WAL inner mutex by offloading file I/O to a dedicated background writer task and enable batching for higher throughput.
• Provide immediate per-record logical IDs at enqueue time and expose clear behavior when the WAL becomes overloaded.

Description

• Replaced the synchronous WalInner/inner mutex model with a bounded tokio::mpsc channel and a WalWriter background task that owns the file and performs batched writes and fsyncs according to WalConfig.
• Reserve logical IDs at enqueue time using an AtomicU64 (next_id) so append returns the assigned id immediately and append enqueues a WalWriteRequest via try_send, returning a WalError::Backpressure or WalError::WriterStopped on overload/stop.
• Moved the in-memory index to Arc<Mutex<HashMap<u64, u64>>> and update it from the writer after each batch; flush is implemented by sending a WalMessage::Flush with a oneshot responder to the writer.
• Moved append_count and bytes_written to atomics updated by the writer, added channel_capacity to WalConfig, and added error variants Backpressure, WriterStopped, and InvalidConfig.

Testing

• No automated tests were executed on this change (existing unit tests remain in wal::tests but were not run as part of this rollout).

---