MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions