Integrate vllm/ns into llm-on-ray

[jiafuzha]

This PR is to replace the closed PR, #264, which is from old branch. This PR merged some enhancements from NS main branch.

1. reshaped neural-speed as a full functional inference engine for vllm
2. integrated vllm ns extension into llm-on-ray and optimized deployment with ray
3. optimized neural-speed in several places, including compute graph construction, multiple numa node deployment and enabling flash attention kernel on llama-3-8b.
4. updated and fixed some benchmark script for IDC test and open-ai mode test, including multiple messages with different roles, removing empty chunk, fixing wrong first token latency and next token latency in open-ai mode.
5. only Llama-2-7b-chat-hf and Llama-3-8b-instruct are supported. But it can quickly extend to support other models.
6. addressed some review comments in last closed PR.
7. 2X perf improvement compared to plain vLLM cpu.