Fix compute_loss signature compatibility with transformers >= 4.46

[ma-kjh]

What does this PR do?

Problem

The latest version of transformers (v4.46+) passes a new argument num_items_in_batch to the compute_loss method.
Currently, the custom trainers (GradAscent, etc.) do not accept arbitrary keyword arguments, causing a TypeError when running with recent transformers versions.

Error log:
TypeError: GradAscent.compute_loss() got an unexpected keyword argument 'num_items_in_batch'

Solution

Added **kwargs to the compute_loss method signature in custom trainers to ensure compatibility with the latest transformers library.

Fixes # (issue)
N/A

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Have you gone through the contributions guide?
• Are your changes documented? Read documentation guidelines here.

[filyp]

I'm also going to need more recent transformers, so I'm wondering why did you close this PR? Was there something wrong with this fix?