[core] separate builder init and builder prepare for each batch

[youkaichao]

Right now we create the builder instance for every batch, and hence create the attention metadata builder instance for every batch, but we don't have some global information when we create the input for every batch.

When upgrading to flashinfer 0.2, see #11194 , the attention metadata builder needs to access some global information such as sliding window, for the plan function before the flashinfer wrapper can be used.

This PR fixes the problem, by separating the builder init and builder prepare for each batch:

• we can remember global config when we create the builder.
• the builder will call prepare (or we can rename it to reset) to prepare for the current batch, and now it can access the global config stored during construction.

[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 starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

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

🚀

[comaniac]

cc @elfiegg @pavanimajety for FlashInfer 2.0 upgrade

[pavanimajety]

Thanks Kaichao! Nice change, LGTM.

[pavanimajety]

We probably need sm_scale too, but can go with the Flashinfer PR.

[youkaichao]

yep, this PR just makes it easier to add more values in the builder. we can add them in the flashinfer pr.