Humans and Deep Networks Largely Agree on Which Kinds of Variation Make Object Recognition Harder