Byzantine-Robust Federated Learning with Optimal Statistical Rates