Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback