Diverging Preferences: When do Annotators Disagree and do Models Know?