GitLab

* Fix tokenization issue in BaseLM.loglikelihood

* Add a regression script

* Use entire non-continuation length as context

---------
Co-authored-by: Hailey Schoelkopf <65563625+haileyschoelkopf@users.noreply.github.com>