Routing to the Expert: Efficient Reward-guided Ensemble of Large Language Models