Mitigating Hallucination in Multimodal Large Language Model via Hallucination-targeted Direct Preference Optimization