Add support to redact model outputs

[MiguelAFH]

As part of MedHELM, we need the ability to redact not only the requests, but also the generated responses. This is to prevent PIH data from real patients.

[yifanmai]

Overall it looks good to me. Couple of notes:

1. Could you run the linter? Install the [dev] extras (or you can install the linter directly using pip install black==24.3.0 flake8==5.0.4 mypy==1.5.1) and then run ./pre-commit.sh, and then make the requested fixes.
2. I would prefer to replace tokens with the empty array [] in case the length of the array leaks some information.