MoColl: Agent-Based Specific and General Model Collaboration for Image Captioning