Separate train script into one to train directly one classifier and one to train one binary classifier (accept/reject) for each discursive function

Add support for labeled data in input tsv and make the paralallel between conversions to and from prodigy's jsonl more obvious

Temporary hack to be able to extract paragraph subcorpora as well article ones (TODO: expose something similar to readNamedTsv returning Either (Vector ParagraphRecord) (Vector ArticleRecord))

Add a script to explode (Text) articles into paragraphs, outputing a files.tsv with a primary key for each of them in the process