Fix: prepare works even if nothing except tp specified (rare)

[S1ro1]

If we only had TP specified, the training would probably be still correct, though the optimizer/dataloader wouldn't be accelerator prepared (in some cases). This was and is a rare case, as pure TP is not fully supported yet anyway. Though this PR fixes atleast this.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

LGMT. Just to be sure, this should be safe wrt to fsdpv2 for example ? Do we handle the states differently for these cases ? If tp is used, this will always to executed first

[SunMarc]

TP training alone should work no ?

[S1ro1]

TP training alone should work no ?

Yes it should, though there are some limitations which I didn't have time to address yet (saving etc), though usage of tp only training is very limited so not a big deal IMO.

LGMT. Just to be sure, this should be safe wrt to fsdpv2 for example ? Do we handle the states
differently for these cases ? If tp is used, this will always to executed first

No, this would only cause _prepare_one to be called multiple times which becomes a no-op the 2nd time anyway.