[ci] refactor eval into api eval and add h800 eval workflow

[zhulinJulia24]

including #3987

1. Refactored chat model evaluation to api-eval
2. Added h800 chat and base model evaluation
3. Added more models to base model evaluation
4. Changed nativa to cuda-ipc

In a future pull request, we will add the openai_extra_kwargs configuration extension (which will add reasoning_effort to gpt-oss) and llmjudger to api-eval.