Support arbitrary processor

[ArthurZucker]

What does this PR do?

Saves the tokenizer class, making sure that the defaults that are stored in transformers are not blocking people from re-using pre-processors. Superseed #30030

For previous models:

import requests
from PIL import Image
from transformers import LlavaProcessor, LlavaForConditionalGeneration

checkpoint = "Intel/llava-gemma-2b"
processor = LlavaProcessor.from_pretrained(checkpoint, tokenizer_class="GemmaTokenizerFast")

But once saved, the tokenizer_class will be part of the processor_config.json, no need to specify it further.

Before merging:

• Add tests
• Add doc

[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.

[amyeroberts]

Nice - thanks for enabling this!

[muhark]

Thank you for adding this and I am so sorry for not following up!