Prune at the Clients, Not the Server: Accelerated Sparse Training in Federated Learning