Making Retrieval-Augmented Language Models Robust to Irrelevant Context