Clustered Sampling: Low-Variance and Improved Representativity for Clients Selection in Federated Learning