Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization