Minimax Optimal $Q$ Learning with Nearest Neighbors