Nearly Minimax Optimal Reinforcement Learning for Linear Mixture Markov Decision Processes