[Bugfix] Fix issues for Pixtral-Large-Instruct-2411

[ywang96]

Previously the mistral-format Pixtral implementation assumes all models under PixtralForConditionalGeneration have the same token id for [IMG_BREAK] and [IMG_END] tokens. This is in fact not true and thus Pixtral-Large-Instruct-2411 is broken on 0.6.5.

This PR adds a workaround to address this issue as well as fixing a bug where image features are inadvertently moved to cuda devices when they don't need to be.

@patrickvonplaten As I mentioned offline, it would be great if these token IDs can be retrieved from the model config so that we won't need to hardcode them.

[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.

🚀