⏩ Train on completion only

[qgallouedec]

What does this PR do?

Fixes # (issue)

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?
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[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.

[LeonEricsson]

Suggestion to update the DataCollator example to:

    Examples:
    ```python
    >>> from trl import DataCollatorForLanguageModeling
    >>> collator = DataCollatorForLanguageModeling(pad_token_id=0)
    >>> examples = [
    ...     {"input_ids": [1, 2, 3, 4], "completion_mask": [0, 0, 1, 1]},
    ...     {"input_ids": [5, 6, 7], "completion_mask": [0, 1, 1]}
    ... ]
    >>> collator(examples)
    {'input_ids': tensor([[1, 2, 3, 4],
                          [5, 6, 7, 0]]),
     'attention_mask': tensor([[1, 1, 1, 1],
                               [1, 1, 1, 0]]),
     'labels': tensor([[-100, -100,    3,    4],
                       [ -100,    6,    7, -100]])}
    ```

given that completion_only_loss is true by default

[qgallouedec]

This is not related to the core change of this PR.
With the new serialisation logic of TrainingArguments, passing a wrong dtype fails when you instantiate the TrainingArguments. There is no need for such test anymore

[qgallouedec]

This concatenation needs to be removed, as we loses the information about where the completion starts. This completion is now managed in tokenize.

[qgallouedec]

Suggestion to update the DataCollator example to:

Thanks! Done in 7f7f2a4