[gpt-oss] Harmony changes with container tool support

[lacora]

Followup on the test failure

Purpose

Support system prompt customization via injections
remove developer messages if all tools enabled are builtin tools
support container tool
support context with session id on MCP
support MCP session clean up for container

Test Plan

Tested on oss gpt with evals, sample requests
Jupyter python tool for AIME25 is slightly lower than model card

Test Result

Verfied AIME results

[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 ask your reviewers to trigger select CI tests on top of fastcheck CI.

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.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[mergify]

⚠️ The sha of the head commit of this PR conflicts with #23386. Mergify cannot evaluate rules on this PR. ⚠️

[yeqcharlotte]

no longer needed given #23386 is merged?