docs: connect native Transformers guides to tested CUDA recipe

[LauraGPT]

Summary

Add a concise CUDA section to the shared English/Chinese native Transformers guides. Link the already merged QwenAudio/Fun-ASR#173 canonical recipe and separate GPU requirements; preserve all CPU Python examples and existing routes. Website and GitHub Pages consume these same source documents.

Validation

• Two new documentation assertions failed before the change; 20 native guide tests passed after and on the signed exact commit.
• Complete product-site suite: 518 passed (four workers; one pytest plugin rewrite warning). A prior serial run timed out at 540 seconds and is not counted as passing.
• Two independent builds yielded identical 271-file output, 204 pages validated, archive member/byte SHA256 verified.
• Six local Chromium EN/ZH cases at 320/390/1440 px passed; representative screenshots inspected.
• Independent read-only review found no P1/P2.
• Underlying H100 runtime evidence: eight float32/BF16 cases and three exact CLI GPU/CPU regression cases, pinned native weights and Transformers 5.17.0.

Boundaries and rollback

No accuracy, keyword improvement, capacity, other-GPU, attention-kernel or hosted Space/Colab claim. Chinese recognition error and attention warning remain explicit. No model/package/runtime changes, no PyPI release. Pre-push source bundle, raw test logs and reproducible site artifact retained on ind-gpu8; production remains unchanged until exact-head CI and merge gates pass. Signed commit and DCO.