MoCHA: Advanced Vision-Language Reasoning with MoE Connector and Hierarchical Group Attention