[serve][llm] Add direct streaming target for LLM tests

[eicherseiji]

Description

Adds a direct-streaming variant of the existing Ray LLM OpenAI compatibility test target.

• Duplicates serve/gpu/integration/test_openai_compatibility.py at the Bazel level with a _direct_streaming target suffix.
• Sets RAY_SERVE_LLM_ENABLE_DIRECT_STREAMING=1 only on the duplicated target.
• Reuses the same Python test file for the normal and direct-streaming runs.
• Skips the existing multi-model no-model-parameter case when direct streaming is enabled, since direct streaming currently supports exactly one LLM config.
• Adds minimal direct ASGI support to MockVLLMEngine so the existing OpenAI compatibility tests can run through the direct-streaming path without a real vLLM frontend.

Related issues

Related to #63167.

Additional information

Validation run locally:

python python/ray/setup-dev.py -y -a llm
python -m py_compile \
  python/ray/llm/tests/serve/gpu/integration/test_openai_compatibility.py \
  python/ray/llm/tests/serve/mocks/mock_vllm_engine.py
python -m pytest --collect-only -q \
  python/ray/llm/tests/serve/gpu/integration/test_openai_compatibility.py
uvx pre-commit run --files \
  python/ray/llm/tests/BUILD.bazel \
  python/ray/llm/tests/serve/gpu/integration/test_openai_compatibility.py \
  python/ray/llm/tests/serve/mocks/mock_vllm_engine.py
git diff --check

The duplicated Bazel target should be exercised in Buildkite.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[eicherseiji]

Closing as superseded: the direct-streaming OpenAI compatibility target has been folded into #63468.