What Do Speech Foundation Models Not Learn About Speech?