Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decoding