A Unified Hallucination Mitigation Framework for Large Vision-Language Models