docs: clarify long-audio segmentation, memory and acceptance

[LauraGPT]

Summary

• Correct both long-audio articles, their search metadata and the matching blog index cards while preserving URLs and unrelated cards.
• Explain Whisper's internal decoding windows rather than a false 30-second file cap; describe full-waveform CPU loading and actual VAD batching/CPU behavior.
• Provide the same explicitly CPU-based SenseVoice/FSMN-VAD Python recipe in both languages, with environment and empty-output checks.
• Replace unsupported speed/coverage promotion with an application acceptance checklist and maintained model/deployment links, including third-party MOSS boundaries.

Verification

• Test-first source/built contracts: original HTML failed; final focused suite passed 59 tests, including controlled execution of the documented code. Controlled model/metadata boundaries do not establish acoustic correctness or installation.
• Full product-site suite: 348 pytest tests across three shards with exact collected-node coverage; the earlier 560-second executor timeout is preserved and not counted as passing evidence. Chromium: 118 tests, including four added mobile/desktop article checks and screenshot review.
• Two validated static builds: 251 byte-identical files, 188 HTML pages. Unrelated index cards and related-post links retained.
• Independent HTML review found no unresolved issues. Core source files byte-identical to the published FunASR 1.4.15 revision.

Documentation/tests only. No new model inference, long-audio benchmark, peak-memory measurement or PyPI release. The 230 MB example is conditional waveform arithmetic, not measured process memory. Signed commit, DCO and verified rollback backup.