SCAFFLSA: Taming Heterogeneity in Federated Linear Stochastic Approximation and TD Learning