FedLGA: Towards System-Heterogeneity of Federated Learning via Local Gradient Approximation