Your fairness may vary: Group fairness of pretrained language models in toxic text classification