Does Differential Privacy Impact Bias in Pretrained NLP Models?