Device Heterogeneity in Federated Learning: A Superquantile Approach