[Model] Consolidate ViTs attention implementation without mask

[Isotr0py]

• This PR adds a MultiHeadAttention layer to consolidate existing duplicate attention forward implementation in ViTs.
• ViTs that use attention mask are not modified: Qwen2-VL, Pixtral, Mllama.

[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.

🚀

[DarkLight1337]

Unblocked multi-modal tests to see if all of those models still have the correct output.

[DarkLight1337]

No new errors are introduced by this PR (compared to this nightly build), so it should be fine.

[DarkLight1337]

Would be nice if you could fix the InternVL tests here as well though! I've already asked for access to MiniCPMV to resolve the gated repo error.

[Isotr0py]

The failing internvl tests should also be fixed in new commit now.