Advancing Multimodal In-Context Learning in Large Vision-Language Models with Task-aware Demonstrations