REVS: Unlearning Sensitive Information in Language Models via Rank Editing in the Vocabulary Space