Pretrained Transformers Do not Always Improve Robustness