Minimax Optimal Q Learning with Nearest Neighbors