[Bugfix][V1] Re-compute an entire block when fully cache hit

[comaniac]

Previously we re-compute the last token when the prompt is fully cache hit. However, since allocate_slots() calculates number of computed tokens based on the number of computed blocks, the number of computed tokens allocate_slots() inferred is always multiply by the block size. This results in the last block fails to be cached in the right way, and breaks the hash block chain in prefix caching. While this doesn't affect the accuracy, it hurts the cache hit rate a bit.

cc @WoosukKwon

[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.

🚀

[WoosukKwon]

LGTM. Thanks for the fix!