Unmasking Efficiency: Learning Salient Sparse Models in Non-IID Federated Learning