Moss: Proxy Model-based Full-Weight Aggregation in Federated Learning with Heterogeneous Models