Design and add a useful CLI

[lebrice]

We'd need to think about the kinds of commands that people will be running, and what kind of things we can setup for them.

Ideas:

• project sync (sync the codebase across all the clusters so it can be used for jobs)
• project train (...): training run, with a validation set when applicable. No testing.
• project eval (...): training (without a validation set), (or load from a checkpoint) then evaluate on the test set.
  • I'd ideally like to encourage not using the test set for validation whenever possible.
  • Other "good ML practices" should also be encouraged or even enforced, if that's desirable.
• project profile (...) same as train, but shorter, with more profiling tools setup?

Idk, just throwing ideas out there.