Move the unknown token replacement step to download

After #239 (closed) is done, we may have an issue with the unknown token.

Example:

• dataset A has no P occurence in its train/val sets but has some in test set.
• all occurences of P in the test set are replaced during dataset extraction
• dataset B is being merged with dataset A
• dataset B has occurences of P in its train/val sets. The character P should no longer be considered unknown.
• we have no way of reverting that change

That's why the change must be done during download, where the dataset is finite.