Guided decoding

[yunfeng-scale]

Pull Request Summary

Support guided_json, guided_regex, guided_choice from vLLM

Test Plan and Usage Guide

tested within vllm entrypoint and from model engine

curl localhost:5005/predict -d '{"prompt":"hello", "guided_regex":"a beautiful world .*"}'
{"text": "a beautiful world is fuxix", "count_prompt_tokens": 2, "count_output_tokens": 16, "log_probs": null, "tokens": ["a", " ", "b", "ea", "u", "t", "iful", " ", "wo", "r", "l", "d", " is", " f", "ux", "ix"]}