Structured benchmark

[willkurt]

This is a revised version of the old PR that fixes some small issues and adds support + data for all the Mistral evals we did as well.
In this PR we should have:

- README documenting all the steps to reproduce.
- Necessary model handlers and additional code necessary to run the models.
- Modal files for deployment for both structured and unstructured generation.
- Data from all the runs so that re-running is not necessary.
- A few Python notebooks so that visualization and error analysis is easier to perform.

[rlouf]

Is this good for review?