Update transformers version to 4.48

[wizeng23]

Description

Llama vision attention is still broken. Nikolai determined it occurs when gradient checkpointing is enabled, and filed an issue here: huggingface/transformers#36040.

Since this can be worked around, and only affects one family of models, we still want to upgrade the version. 4.45 is 4 months old, and we need newer transformers versions to support new vision models.

We attempted this previously, and reverted it in #1111 due to the same issue.

Main code changes required are:

• Rename tokenizer to processing_class
• Update some attention class names due to HF Transformers' attention refactor 🚨All attention refactor🚨 huggingface/transformers#35235

Related issues

Fixes OPE-738, OPE-875
Towards OPE-1018

Before submitting

• This PR only changes documentation. (You can ignore the following checks in that case)
• Did you read the contributor guideline Pull Request guidelines?
• Did you link the issue(s) related to this PR in the section above?
• Did you add / update tests where needed?

[xrdaukar]

What config was used for Llama Vision testing?

[wizeng23]

configs/recipes/vision/llama3_2_vision/sft/11b_lora/gcp_job.yaml. Added to PR description