new grpc server and client for new knobs.

[ZHANGWENTAI]

New grpc server and client for tuning inference framework, dtype, input length and batch size on LLMs from huggingface.

Ⅰ. What this PR does / why we need it

• can specify LLMs on huggingface and download
• new grpc server for multiple inference frameworks: native torch and vllm
• int8 quantization with bitsandbytes
• new grpc client for stress test under different input lengths and batch sizes
• example for tuning new knobs: inference framework, dtype, input length and batch size

II. Does this pull request fix one issue?

implement the first part of #52