[WIP][Core] Support tensor parallelism with uneven heads

[yixqiao]

Currently working on supporting tensor parallelism when using attention heads that do not divide the number of GPUs evenly, for example with Qwen2.

Using a dummy head approach to pad to an even number. Currently testing with Qwen2 1.5B with 12 heads on 8 GPUs (so 4 dummy heads).

Works end to end, but has poor results.
Prompt: "This is a brief introduction to large language models. Large language models (LLMs) are"
vLLM non-parallel reference: "particularly useful language model techniques to interact with language models, so as to compress a"
New feature (this PR): "10 instructions in deleting newton's law of relational networks. Common utility."

Working on debugging the accuracy issue now.

FIX #11797

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @yixqiao.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork