Model Pruning Enables Efficient Federated Learning on Edge Devices