[BugFix][Diffusion] Scope Wan RMSNorm patch to NPU

[Dong1017]

Fixes #8030.

Importing the Wan 2.2 package replaced diffusers' WanRMS_norm globally, including on CPU and CUDA. The Wan VAE fastpath only supports the original diffusers class, so its normalization optimization was skipped and six tests failed depending on import order. Apply the automatic replacement only on NPU, restore module aliases after tests that explicitly exercise the patch, and correct the moved RMSNormVAE test import.

Validation: reproduced the reported six failures on upstream 75570ee3a with diffusers 0.40.0; both affected test files now pass together (79 tests). ruff check, ruff format --check, and git diff --check pass. The reproduction host has vLLM 0.29.0; the reported CI used 0.30.0.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/offloader.md, docs/design/module/diffusion/diffusion_model_integration.md, docs/design/module/diffusion/index.md.

Module owners: @wtomin @david6666666 @Isotr0py

Routing: @wtomin via module of the changed files, module named in the PR description, CODEOWNERS; @david6666666 via module of the changed files, module named in the PR description; @Isotr0py via module of the changed files, module named in the PR description

@Dong1017, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[Gaohan123]

LGTM. Thanks

[Dong1017]

Self-review (precheck-pr quick, b1bd11f): I checked all three changed files against #8030: the NPU-only patch guard, restoration of WanRMS_norm aliases after patch tests, and the moved RMSNormVAE import. With diffusers 0.40.0 and vLLM 0.29.0, the combined tests reproduced six failures on main and passed 79/79 on this commit; Ruff and diff checks passed. The guard also changes CPU/CUDA runtime normalization selection, so full model-output parity remains unverified. Current NPU CI failures are in MiniCPM-o perf and Qwen3-TTS steps, while AMD and Intel CI are pending. The repository precheck also flagged title prefix casing and a legacy SPDX header for follow-up after CI.

[Dong1017]

CI triage for b1bd11f2a:

• Passed: main Buildkite, Intel, DCO, Python builds, and pre-commit.
• AMD: the hard failure is in the Qwen-Image Custom Pipeline Test. Two cases report Sleep mode can only be used for one instance per process and model parallel state must be empty before initialization.
• NPU: MiniCPM-o 4.5 loses its EngineCore/Stage-0 replica; its perf test has request failures, and Qwen3-TTS also reports no live Stage-0 replica. The underlying cause of the worker exits is not established by the test summaries.

These failing jobs do not run the changed Wan VAE tests. The production guard in this PR still calls patch_wan_rms_norm() on NPU. I found no failure pointing to the changed Wan VAE path, so I have not changed unrelated code. Retrying only the failed Buildkite jobs on this same SHA would help distinguish transient CI failures from reproducible ones.

[Dong1017]

Root-cause timeline for #8030:

• [Perf][Wan2.2] Add fused RMSNorm replace WanRMS_norm on npu #3067 introduced an unconditional patch_wan_rms_norm() call when the Wan2.2 package is imported. Its stated goal and test plan were NPU fused RMSNorm, but the call also replaces WanRMS_norm bindings in already-loaded modules on CPU/CUDA. [BugFix] Probe __dict__ instead of hasattr when patching WanRMS_norm #3400 made that scan safer without changing its platform scope.
• [AutoRound] Support WAN2.2 W4A16 quantization model #3353 added tests that call the patch explicitly but did not restore those bindings afterward. [optimization] Added Wan 2.2 VAE decode optimizations #7056 later added the GB200 Wan VAE fastpath and its CPU tests. That fastpath deliberately recognizes the original diffusers WanRMS_norm, not the substituted RMSNormVAE. When the patch tests run first in the same pytest process, newly constructed VAEs contain the substitute, so RMSNorm fastpath discovery and related assertions fail.
• Separately, [Refactor][Diffusion] Move single-model layers into their model directories #5908 moved RMSNormVAE to models/wan2_2/norm.py; a new [optimization] Added Wan 2.2 VAE decode optimizations #7056 test still imported its old location, causing the independent ImportError.

This PR limits automatic substitution to NPU, restores module bindings after the explicit patch tests, and corrects the import. With diffusers 0.40.0 and vLLM 0.29.0, the combined-order run reproduced six failures on the base revision and passed 79/79 on this revision. The CPU/CUDA class selection changes as a result; complete model-output parity remains unverified.