[Tracking] Improve Multimodal CI coverage

Checklist

• 1. If the issue you raised is not a feature but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 2. Please use English, otherwise it will be closed.

Motivation

1. All existing Multi-modal CI operations are executed on a single GPU, without considering the scenario of Tensor Parallelism (TP). There is a requirement to introduce test cases for TP-2/4 configurations.
2. The payload for VLM CI is relatively low. Stress tests are necessary, and simultaneously, it is crucial to investigate whether there are any memory leaks.
3. At present, the majority of the VLM CI is only to vision. It is essential to expand its scope to audio.
4. Welcome to add more.

Related resources

The test is under test/srt/test_vision_openai_server_x

Related PRs:

• https://github.com/sgl-project/sglang/pull/8428
• https://github.com/sgl-project/sglang/pull/7519