Fix perplexity computation in perplexity.md

[Framartin]

What does this PR do?

This pull request fixes an issue in the code of the Perplexity of fixed-length models documentation. The current implementation computes the arithmetic mean of the negative log-likelihood (NLL) per sequence chunk without accounting for the varying number of tokens in each chunk. This can lead to an inaccurate average NLL per token, as tokens in smaller chunks (like the last one) have a higher weight in the final perplexity calculation. The PR computes the arithmetic mean of NLL per token. The perplexity error is small in the provided example (16.4541 vs. 16.4443), but I guess that it can be larger for shorter sequences.

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?
• 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?

@stevhliu @ArthurZucker

[ArthurZucker]

Thanks

[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.