Support VLM model and GPT4V API

[xunfeng1980]

VLM model: Qwen-VL/LLaVA etc
VLM API: GPT-4V API https://platform.openai.com/docs/guides/vision

[l4b4r4b4b4]

I strongly second this. However Llava 1.6 models are very peculiar :D

[Reichenbachian]

It's fine if it's peculiar. We can just error (BadRequest) anything that doesn't have image first or multiple images for instance, but still keep the same format. We shouldn't design the endpoint around the model. The model should fit into the endpoint.

[ywang96]

Closing this as we merged #5237