save json files for pre-existing predictions

[zenssh]

• if qa.json file, if an entry has key llm_predictions with pre-computed predictions(List[str], same length as entry["conversations"]), it will be written into json file so that inference step can be skipped.
• in this case ntokens and proctime will be nulls (unless they are present in qa.json)
• entry["id"] must be present, because entry["video"] is a potentially nested path, which is not good to use for a file name. Also, extension might not be 4 chars. For youtube videos id will be "<youtube_id>_<segment_num>".