docs: make SenseVoice emotion-tag guidance reproducible and bounded

[LauraGPT]

Summary

• Rewrite the existing English/Chinese SenseVoice emotion-tag article around raw output, display processing and evaluation. Preserve its URL and original publication date.
• Remove unsupported speed/accuracy claims; distinguish predicted tags from human state, identity, calibrated confidence and ground truth.
• Show a real 5.55-second public-sample CPU readout and waveform; keep constructed emoji/WA-UA examples explicitly synthetic.
• Link the maintained installation and pinned SER evaluation guides. Preserve unknown/unparseable cases and the unresolved community reproduction issue.
• Keep the blog homepage's one lead and four selected stories unchanged.

Verification

• Final site suite: 495 passed.
• Six real Chromium cases: EN/ZH at 320/390/1440px; screenshots inspected, images loaded, language navigation works, no page errors/overflow.
• Regression: missing installation prerequisite fails before the fix; copy control geometry fails before the scoped padding fix.
• Exact bilingual article Python snippet extracted and executed on Linux CPU with FunASR 1.4.15 and PyTorch/torchaudio 2.10.0. Only model/input path literals were replaced with the cached official weights/public sample.
• Real postprocessor and maintained SER parser/metric checks passed. This is not an emotion-accuracy benchmark or a reproduction of CASIA/RAVDESS scores.
• Two identical builds: 204 validated HTML pages, 271 files.
• Independent seven-file review: previous installation P2 fixed, no remaining P1/P2.

Publication

Signed and DCO-signed commit. Rollback source bundle, patch and acceptance receipts saved before push. Deployment remains gated on exact-head CI and merge. No runtime changes, package release or issue closure.