Paragraph-based Transformer Pre-training for Multi-Sentence Inference