Bumping transformers and st versions

[NohTow]

This PR bumps transformers and sentence-transformers versions.
Transformers bump allows the use of ModernBERT.
Sentence-transformers allows to fix an issue to load PyLate model trained with trust_remote_code=True that was introduced in a modification of the code of ST back then.

This is a draft because I identified some breaking changes that I'll fix.

[NohTow]

Well it seems that the issue I had identified was the only one (minus the less verbose logging of NanoBEIR now).
The only change I had to make was to make set_widget_examples a no-op because there is issues with it and our data mapping/transform. I already did a modification to skip its call back then, but Tom reintroduced a call to it, so the best thing to do is just make it a no-op for now, except if you want to debug.

Tests are passing and I have been training with transformers 4.48.2 for some time, as well as a new version of ST (not the final one though) and had no issue. As tests are passing, I guess it is safe to merge @raphaelsty

[NohTow]

I fixed a small issue where the model cards created when training on top of a ST models were just copy of the existing model card instead of the new one, due to a change in ST.
I actually removed the overriding of the create_model_card function because it was useless and was not updated due to this.
Maybe at some point it would be great to do a cleaning pass a remove any remaining useless overriding to better profit from ST changes.
Other changes related to the removing of the Trainer in the model card seems to be handled correctly because we extend the objects correctly.