BCOTrainer version upgrade fixes

[claralp]

What does this PR do?

This PR fixes to some issues that I identified with the BCOTrainer on the latest version upgrade:

• remove extra shuffles, super class already handles data shuffling. These destroy training reproducibility
• fix serialization of UDM classifier, not all parameters were saved previously
• accelerate.gather not hanging on empty lists anymore
• convert delta to tensor before logsigmoid, this led to numeric errors
• do not update delta value in eval mode

Let me know if you have any questions

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines. - changes do not affect docs
• Did you write any new necessary tests? - tests exist already

Who can review?

@kashif @qgallouedec

[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.

[kashif]

@claralp we will need to add joblib as a dependency?

[claralp]

@kashif yes, but could use pickle as alternative if you prefer. Just BCO with UDM needs this

[PhilipMay]

Well, IMO using joblib is a good idea because it supports pickle with compression with just a line of code. So complexity is reduced. But it can also be a good idea to reduce the number of external dependencies. On the other side joblib is a very common dependency.

So @kashif should decide this. What do you think?