Distinguishing the Knowable from the Unknowable with Language Models