Dynamic Adaptive Shared Experts with Grouped Multi-Head Attention Mixture of Experts