Extracting Unlearned Information from LLMs with Activation Steering