feat: serve MOSS diarization across deployment stack

[LauraGPT]

Summary

• expose the third-party OpenMOSS MOSS-Transcribe-Diarize model as a first-class funasr-server alias with a pinned HF revision and native anonymous speaker segments
• add dedicated GPU Docker Compose and Kubernetes recipes without attaching an external VAD or speaker model
• publish one consistent service contract across the four READMEs, OpenAI client/agent docs, GitHub Pages source, deployment matrices, FunClip links, and the bilingual funasr.com deployment hub

Runtime boundaries

• MOSS remains an OpenMOSS Apache-2.0 model; FunASR owns the adapter and service/documentation contract
• the FunASR HTTP path is offline file transcription, not realtime WebSocket serving
• spk=true does not start a second diarization pipeline for MOSS; verbose_json preserves the native S01/S02 labels
• vLLM, SGLang Omni, and LocalAI remain their respective native service paths; the docs do not claim identical output across runtimes

Exact-head verification

Head: e7ddad713ed056e553e920c1c4a33adf05dc5bb6

• 151 passed across the MOSS HF/vLLM/SGLang adapter, built-in HTTP service, documentation contracts, and full product-site Python suite
• real H100 80GB HTTP smoke with Transformers 5.16.0.dev0, Torch 2.11.0+cu130, and pinned model e8681d68: the 6.0 s repository sample returned HTTP 200, duration=6.0, one non-empty monotonic segment, and speaker S01
• product site built 29 generated pages and validated 110 total pages
• Playwright: 4 MOSS/FunClip mobile and desktop tests passed; desktop and mobile screenshots inspected
• Python syntax, OpenAPI JSON, Kubernetes YAML, Black-managed files, and git diff --check passed

The dedicated container files were statically validated, but ind-gpu8 has no Docker CLI, so this PR does not claim an image-build result. The same source path was validated through the real GPU HTTP service.

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com