Text Embeddings by Weakly-Supervised Contrastive Pre-training