Erasing Conceptual Knowledge from Language Models