Mitigating Object Hallucinations in Large Vision-Language Models via Attention Calibration