Support decode token logprobs

[comaniac]

Initial support to logprobs of decoding tokens:

• When return_logprobs is set, the response includes an additional field token_logprobs, which has a list of logprob for each token including prompt tokens. Each element in this list contains (token text, token ID, logprob).
• When streaming, the token_logprobs behaves as text. In other words, it is monotonic.
• This PR only supports top 1 logprob.
• For OpenAI APIs, text_offset and top_logprobs are not supported yet.

Known issue:

1. The request with regex + streaming + logprobs, the logprobs is not monotonic due to regex's decoding process. Since this kind of requests doesn't make sense, we won't fix this issue for now.
2. To align with OpenAI logprob style, there are additional overheads such as tokenizer.convert_ids_to_tokens.

cc @merrymercy @Ying1123 @Ja1Zhou

[comaniac]

@merrymercy comments addressed. I've verified the performance of benchmarking Hellaswag. The latencies before and after this PR are almost the same.