Fixing Imbalanced Attention to Mitigate In-Context Hallucination of Large Vision-Language Model