General Purpose Text Embeddings from Pre-trained Language Models for Scalable Inference