fix loss scaling only when compute_loss_func is used

[BlackNoodle]

What does this PR do?

In #34198, the line loss *= self.args.gradient_accumulation_steps was introduced due to Negate accelerate grad accum div. This change was made to correct errors encountered during gradient accumulation. However, the scaling should only occur when compute_loss_func is used, so the code was modified accordingly.

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 or the forum? 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, and
  here are tips on formatting docstrings.
• 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.
@muellerzr
@ArthurZucker

[ArthurZucker]

cc @muellerzr compute_loss_func is never None right?

[muellerzr]

No, it should always be used, period. The correct check is if we passed in num_items_in_batch or not. If not, then we can mult.

(There are model loss functions which accept num_items_in_batch as part of their forward)

[muellerzr]

This will conflict with #34373, however the correct version of this is either if compute_loss_func is used or self.model_accepts_loss_kwargs is True