Visual Anchors Are Strong Information Aggregators For Multimodal Large Language Model