test(e2e): add Qwen3 MoE CUDA coverage

[sinpeyw]

Purpose

Add repository-owned CUDA E2E coverage for Qwen3 MoE after #216, following the E2E contract introduced by #222 and refined by #238.

Changes

• Add Qwen3-30B-A3B coverage for the four default E2E scenarios: baseline-graph, afd-eager, afd-graph, and afd-graph-dbo.
• Extract the shared environment, device, and runner-subprocess logic required by Refine test cases #238 before adding a second model.
• Update the E2E documentation and run-e2e skill for the Qwen3 MoE suite.

The production code, shared E2E runner, CI configuration, and DeepSeek-V2-Lite scenario behavior are unchanged.

Validation

• pre-commit run --files ...
• python -m pytest -q tests/unit: 571 passed, 46 skipped
• 5 DeepSeek-V2-Lite parameterized cases, 1 async CAM smoke, and 4 Qwen3 MoE cases collected
• Default Qwen3 GSM8K-7 CUDA gate on NVIDIA A800 GPUs: 4 passed, 0 skipped
• All four scenarios evaluated 7 samples with 1.0 strict and flexible exact match

Coordination

Rebased onto main after #238.

Related to #215, #216, #222, and #238.

[jiangkuaixue123]

@yujuancao07 Please help review this.

[hsliuustc0106]

[P2] Use Qwen node IDs in the run-e2e command

At .agents/skills/run-e2e/SKILL.md:101-102, the instruction says to replace only the test path in the preceding command. That command still contains ::test_deepseek_v2_lite[...]; after replacing the path, pytest cannot collect the Qwen cases because the new function is test_qwen3_moe. Please provide a complete Qwen command, such as python -m pytest -q -s tests/e2e/models/qwen3_moe/test_qwen3_moe.py, or list the four Qwen node IDs.

Validation: full tests/unit passed; the Qwen model unit tests passed under vLLM 0.26; Ruff and compile checks passed. CPU-only CI is green; GPU/NPU E2E was not run locally.

[sinpeyw]

[P2] Use Qwen node IDs in the run-e2e command

At .agents/skills/run-e2e/SKILL.md:101-102, the instruction says to replace only the test path in the preceding command. That command still contains ::test_deepseek_v2_lite[...]; after replacing the path, pytest cannot collect the Qwen cases because the new function is test_qwen3_moe. Please provide a complete Qwen command, such as python -m pytest -q -s tests/e2e/models/qwen3_moe/test_qwen3_moe.py, or list the four Qwen node IDs.

Validation: full tests/unit passed; the Qwen model unit tests passed under vLLM 0.26; Ruff and compile checks passed. CPU-only CI is green; GPU/NPU E2E was not run locally.

The Qwen3 section now provides a complete pytest command.