Compressed and distributed least-squares regression: convergence rates with applications to Federated Learning