feat: support iteration level scheduling

[kemingy]

          Also https://www.usenix.org/conference/osdi22/presentation/yu

Originally posted by @VoVAllen in #382 (comment)

[kemingy]

Although Orca coupled the scheduler and execution engine, it still has something we can learn from.

For GPT-like models, they can benefit from iteration-level scheduling in the following part:

1. status request can return to the client before other requests are finished in this batch
2. new requests can enter the batch without waiting for all the requests in the previous batch to have been finished

---

refer to:

• Orca: https://www.usenix.org/conference/osdi22/presentation/yu
• BatchMaker: https://cs.nyu.edu/~lingfan/resources/batchmaker-eurosys18.pptx
• text-generation-inference continuous batching: https://github.com/huggingface/text-generation-inference/blob/main/router/README.md#continuous-batching