[Usage]: Passing a guided_json in offline inference

### Your current environment

```text
vllm 0.4.2
```


### How would you like to use vllm

I'm trying to force a json generation using outlines in offline inference. I don't see anything related in the documentation.

I haven't found an example of chat completion for offline inference, but I've managed to mimic it using chat templates, this is why I need to force a json generation.
