Compare NLLB and MADLAD-400

[isaac091]

Use the set of experiments provided by the projects team to compare the two models.

We'll want to test single-GPU (using LoRA for MADLAD) and multi-GPU performance (NLLB 3.3B and full MADLAD).

[ddaspit]

@isaac091 Have we tried full-finetuning of MADLAD yet? We now have the H100s, so it should be doable.

[isaac091]

We haven't. I'm interested to see how it does since I don't feel like we have a good idea of what it is or isn't capable of yet.