Do Large Language Models Latently Perform Multi-Hop Reasoning?