Which one makes more sense in terms of scalability and a lot of different versions to exist?
Option B. Here you can read / write single document versions without loading the whole JSON string (and then object, when it comes to manipulation) into the memory. Especially when it comes to many and huge documents, this will save you some time.
However, if you have lots of versions of one huge document that do not differ much from each other, it might be better in terms of performance and memory-usage to store them with incremental or differential versioning, which would make more sense in one single "JSON file".