Byzantine-Robust Distributed Learning: Towards Optimal Statistical Rates