FedSAC: Dynamic Submodel Allocation for Collaborative Fairness in Federated Learning