Minimax Optimal Online Imitation Learning via Replay Estimation