Don't always re-train when creating the documentation

[johannes-spies]

Right now, tox -e docs runs some training steps, which can take a long time. This PR adds a switch to skip training and directly builds the docs with TRAIN=0 tox -e docs.

---

📚 Documentation preview 📚: https://metatrain--797.org.readthedocs.build/en/797/

[Luthaf]

We should merge this together with #786 which is doing something similar for the tests. Ping @HaoZeke =)

[PicoCentauri]

Is this still necessary?

With #854 most of the trainings gets launched inside the example and will be cached. The only one left is the bash example and this is rather quick.

[johannes-spies]

If they're cached, I don't see a problem and this PR here can be closed without merging. But does training happen at least once while building the documentation?

[PicoCentauri]

Yes, when you have no docs, sphinx gallery generates the files and while doing it will call the mtt to perform training via a subprocess