Support Muon, stop using Dataloader2

[norabelrose]

This PR does two main things:

1. We don't use Dataloader2 because that's not supported anymore
2. We add support for the Muon optimizer, because we've found that it dramatically speeds up convergence and leads to much "better" (in the sense of KL loss) lenses

There are also a handful of smaller changes, like using a sample from the SmolLM2 corpus as the default dataset, rather than the Pile, since the Pile is not actually on the HF Hub anymore.