Trusting RoBERTa over BERT: Insights from CheckListing the Natural Language Inference Task