A Joint Approach to Local Updating and Gradient Compression for Efficient Asynchronous Federated Learning