[Core] Factor out input preprocessing to a separate class

[DarkLight1337]

Following up to #7258, this PR implements the suggestion by @robertgshaw2-neuralmagic to create a new class for prompt parsing. This helps keep the sync/async versions of the preprocessing code together so they can be updated together easily, while also slimming down the existing engine classes to focus on request processing.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which consists a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of default ones by unblocking the steps in your fast-check build on Buildkite UI.

Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge).

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[njhill]

Thanks @DarkLight1337, just some minor comments.

We may need to re-think some of this soon since we're hoping to get rid of AsyncLLMEngine altogether. I'm thinking we'll hopefully no longer need both sync and async versions of these things.

[njhill]

have a prompt_type = parsed["type"] and reuse instead of repeated lookups? (same for parsed["content"])

[DarkLight1337]

I tried this, and unfortunately this breaks the type checker's capability to narrow down the type of parsed. So I'm keeping the code as is.

[njhill]

have a prompt_type = parsed["type"] and reuse instead of repeated lookups? (same for parsed["content"])

[DarkLight1337]

As above.

[DarkLight1337]

Thanks for the review! I'll address the detailed comments shortly.

We may need to re-think some of this soon since we're hoping to get rid of AsyncLLMEngine altogether. I'm thinking we'll hopefully no longer need both sync and async versions of these things.

Even without AsyncLLMEngine, the preprocessing logic better fits inside vllm.inputs module, and this PR would reduce the bloat inside LLMEngine itself.

[DarkLight1337]

I have finished addressing your other comments.