Certifiably-Robust Federated Adversarial Learning via Randomized Smoothing