Transformers serve fix

[LysandreJik]

Fix a few issues with transformers serve:

• Some models don't have EOS tokens; on these, we cannot check if that token is generated and purely rely on the length for the length cutoff.
• Some models (like Gemma3) load an image processor with AutoProcessor even if the image processor cannot be loaded. For these, we try/except and load with a tokenizer if we fail.

[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.