SpaFL: Communication-Efficient Federated Learning with Sparse Models and Low Computational Overhead