[Test] diffusion: Cover CUDA and HIP latency guard policies

[Tokha233]

Motivation

Diffusion latency guards deliberately reject over-budget CUDA runs but only warn on HIP. Tests must cover both policies without depending on the host GPU.

Upstream #42391 already merged the fixture platform isolation originally proposed here. This PR is now limited to four additional regression cases: E2E/load latency crossed with CUDA/HIP policy. It checks that an in-budget request emits no warning, an over-budget CUDA request raises, and an over-budget HIP request emits exactly one AMD performance warning. Production validators and thresholds are unchanged.

Validation

Synced with upstream bab04cd7913ff50d1bf1a677a2d10e5015dc4c3b. The remaining diff is 21 lines in test_load_inclusive_e2e.py.

RTX 5090 D v2, Python 3.12, PyTorch 2.13.0+cu130:

• Native CUDA platform policy: 95 passed.
• Simulated HIP platform policy on the same host: 95 passed. This tests policy dispatch, not AMD kernels.
• Changed-file pre-commit and git diff --check passed.
• Remote source content was checked against the complete Git index plus working-tree changes before testing.

python -m pytest -q \
  python/sglang/multimodal_gen/test/unit/test_load_inclusive_e2e.py \
  python/sglang/multimodal_gen/test/unit/test_runner_perf_baselines.py \
  python/sglang/multimodal_gen/test/unit/test_sequential_server_requests.py

No inference speedup or native AMD execution is claimed. Codex assisted with implementation and testing.

---

CI States

Latest PR Test (Base): ❌ Run #37233246273
Latest PR Test (Extra): ❌ Run #37233246161
Latest PR Test (AMD ROCm 10): ❌ Run #37233246322

[Tokha233]

Resolved the upstream conflict in 7c892598. #42391 already contains the platform fixture fix; this PR now contributes only four E2E/load × CUDA/HIP policy regression cases. Fresh validation: 95 passed with native CUDA policy and 95 passed with simulated HIP policy. Production thresholds are unchanged. Local pre-commit and GitHub lint passed.

Logs and exact diff: https://github.com/Tokha233/ComfyUI-H3-SpeedKit/tree/8791760/evidence/pr-maintenance-1005 . The new GPU workflows stop at the run-ci authorization gate; maintainer authorization is needed for hosted validation.