[WiP] Add buf_init_callbacks for dist inference = auto generate rope embeddings after meta/fake tracing

[lessw2020]

add buf_init_callback.
This is an extension of the modeling_llama.py PR here:
https://github.com/huggingface/transformers/pull/32428/files

What does this PR do?

This adds the same class level buf_init_callback dict that allows someone to regenerate the rope embeddings that are destroyed during meta and then fake tracing. (meta can be resolved via intercepting buffers, but running a fake tensor to trace for PP then destroys the rope embeddings as they become fake tensors.

The buf_init_callbacks allows you to quickly find the entry point to regenerate all the non-persistent embeddings after tracing ensuring you have a ready to run model graph.

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 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.

[amyeroberts]

cc @ArthurZucker @gante

[ArthurZucker]

Answered on #32428