Directionality and representativeness are differentiable components of stereotypes in large language models