Mitigating Hallucinated Translations in Large Language Models with Hallucination-focused Preference Optimization