[MATH] Fix generation for chat models & fix normalization for predictions (#163)
* Fix _fix_fracs in MATH normalization
* Fix generation for chat models
* Use same normaliser for predictions and golds
* better stop token managment when using chat templates
* fix test, separate math and math cot
---------
Co-authored-by: Nathan Habib <nathan.habib@huggingface.com>
Co-authored-by: Clémentine Fourrier <22726840+clefourrier@users.noreply.github.com>
Co-authored-by: clementine@huggingface.co <clementine@huggingface.co>