[CB] Update spyre model runner for new spyre input batch

[wallashss]

This PR primarily integrate the model runner with the new spyre input batch for continuous batching.

Summary:

• Removed some hardcoded code to related to sampling parameters. Now, it should work fine requests with sampling parameters with CB.
• Update the logic in warm up to properly clean the data of the model runner, so it can have a consistent state to receive new requests.
• Now we have two input batches: one for the decoding and other for the prefill. Currently, we can only prefill one request at a time, and the batch is reset for every new prefill. This is similar to how the static batching works. This solution also ease to split the requests, because we only have to swap the batch to process without the need to activate/deactivate requests that are in decoding step.

For #76

[github-actions]

👋 Hi! Thank you for contributing to vLLM support on Spyre.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, first install the linting requirements, then run format.sh and commit the changes. This can be done with uv directly:

uv sync --frozen --group lint

Or this can be done with pip:

uv pip compile --group lint > requirements-lint.txt
pip install -r requirements-lint.txt
bash format.sh

Now you are good to go 🚀

[wallashss]

This should be follow up for #126 and #107

[wallashss]

Currently, it pass in the tests of #79

[wallashss]

@yannicks1, considering that the previous execute_model should correctly clean up the request of the model runner do we still need to reset the tkv?

[yannicks1]

considering that the previous execute_model should correctly clean up the request of the model runner

does it reset the tkv too? @wallashss

[wallashss]

Currently, no. I kept the reset by the worker. I should be fine.

[prashantgupta24]

lgtm!