Fix SetFitModel: not a dataclass, not a PyTorchModelHubMixin

[Wauplin]

Related to discussion triggered by huggingface/huggingface_hub#2157 (also internal slack).

SetFitModel class is currently a dataclass which doesn't seem necessary and can have weird side-effect (especially with ModelHubMixin). This PR:

• removes @dataclass from SetFitModel
• makes SetFitModel inherit from ModelHubMixin instead of PytorchModelHubMixin. The former is the generic class. The later is a specific class for model inheriting from nn.Module which is not the case here. The only benefit of PytorchModelHubMixin is to have _from_pretrained and _save_pretrained already implemented to load torch weights but in SetFit those methods are inherited anyway (since we don't want to load torch weights)
• adapts SpanSetFitModel and PolarityModel accordingly
• (updates a type annotation List[str] -> Iterable[str] because it's a generator)

cc @lewtun @tomaarsen @LysandreJik changes should work correctly but I haven't tested them extensively (especially the SpanSetFitModel subclass). Please let me know if you think of any oversight/other things to change.

[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.