feat(moss): add SGLang AutoModel backend

[LauraGPT]

Summary

• add backend="sglang" for the third-party OpenMOSS MOSS-Transcribe-Diarize model through the OpenAI-compatible transcription endpoint
• request verbose_json with max_new_tokens and normalize validated speaker segments into FunASR text, raw_text, timestamp, and sentence_info
• reject malformed labels, decreasing start times, and SGLang fallback S01 labels that are not backed by the raw model transcript
• preserve legal overlapping speech and SGLang timestamp clamping
• update bilingual guides, four deployment matrices, and the product deployment registry with the exact SGLang source pin

The model and weights remain published and maintained by OpenMOSS under Apache-2.0. FunASR provides only the ecosystem adapter and deployment contract. External vad_model and spk_model remain intentionally unsupported for this end-to-end model.

Validation

• exact signed+DCO head: 783c24a
• 40 model, documentation, and product-registry tests passed
• MOSS product-page render test passed
• Black, JSON validation, compileall, and git diff check passed
• independent review found no remaining Critical or Important issues

Runtime evidence boundary

No live SGLang worker was claimed in this change. The available H100 host has a CUDA 12.8 driver while the pinned current SGLang Omni path requires CUDA 13. The adapter is contract-tested against the exact public API/source behavior at SGLang Omni 3f819f9cdae3d4eeec22f73306c9067a1ec2542e. The published #914 H100 benchmark remains upstream runtime evidence only, not a FunASR adapter benchmark.

[LauraGPT]

CI follow-up: the first contract run failed only its Black check. I reproduced the exact formatter locally with Black 25.1.0, confirmed the diff was two mechanical line wraps, amended the signed+DCO commit, and reran Black plus all 40 focused tests and the product-page render test before pushing exact head 783c24a.