DA-MoE: Towards Dynamic Expert Allocation for Mixture-of-Experts Models