Boosting Robustness by Clipping Gradients in Distributed Learning