fix(qwen3.5): allow PP+MTP on disaggregated prefill

[nvpohanh]

[by Codex]

Motivation

Qwen3.5 gained the pipeline-parallel speculative-prefill runtime in #35758, but two older assertions still reject the supported configuration. The published PP4 AgentX recipes therefore have to set PYTHONOPTIMIZE=1 to start.

This removes that workaround for the supported path without opening unsupported PP+spec combinations.

Changes

• Allow resolved EAGLE/NEXTN with PP only for Qwen3.5 architectures, non-overlap scheduling, and disaggregation-mode=prefill.
• Keep decode, non-disaggregated, multi-layer EAGLE, other algorithms, and other model architectures rejected.
• Exempt Qwen3.5 stage-local target layers from the older generic PP/MTP layer-count assertion.
• Add server-argument and layer-compatibility regression tests.

Related: #39378 handles the corresponding DeepSeek/GLM path independently.

Validation

• BLACK_NUM_WORKERS=1 SKIP=no-commit-to-branch pre-commit run --all-files --show-diff-on-failure
• Focused layer compatibility tests: 2 passed.

---

CI States

Latest PR Test (Base): ❌ Run #34967629389
Latest PR Test (Extra): ❌ Run #34967629188
Latest PR Test (AMD ROCm 10): ❌ Run #34967629495

[nvpohanh]

cc @YAMY1234

[YAMY1234]

Thanks for the fix. This overlaps almost entirely with #39378 : applying both yields conflicts in validation_hook.py, layer_setup.py and test_server_args.py, and the layer_setup change becomes dead code once #39378 removes the assertion. Suggest closing this PR and adding the four Qwen3_5* architectures to _PP_EAGLE_SUPPORTED_ARCHITECTURES in #39378?