Scalable Extraction of Training Data from (Production) Language Models