RAFT: Support node restarts.

- Implement a persistent WAL based on RocksDB. We use RocksDB because
given a RAFT entry at index i, all the entries which exist after i need
to be deleted. So, RocksDB became a natural choice to write such a
thing.

- Also, each server can run multiple RAFT groups, and we want to store
them all in one persistent storage.

- Batch all messages to be sent to other nodes, with a small timeout of
10ms. Ignore any nodes which are down and can't receive messages, or any
other errors. RAFT would automatically retry those.

- Remove code from cluster and commit package.