docs: clarify SRT and WebVTT subtitle export workflow

[LauraGPT]

Summary

Rework the existing English and Chinese subtitle tutorial around one reproducible Mandarin workflow: Paraformer on CPU -> SRT -> FFmpeg WebVTT conversion -> manual review.

Keep the published URLs and original date. Show actual short-sample outputs, distinguish export success from alignment quality, and explain that the CLI has no VTT output option or automatic speaker-name prefixes. Keep native Transformers text-only boundaries explicit. The homepage remains one lead and four selected stories; no new article or runtime change.

Validation

• 502 product-site tests passed on the final candidate with the existing four-worker configuration.
• Six browser cases passed (EN/ZH, 320/390/1440): discovery, language switching, waveform, overflow and copy controls. Six representative screenshots inspected; final normalized builds are byte-identical.
• Exact bilingual shell commands ran using the published FunASR 1.4.15 CPU package. Its CLI bytes match source.
• Synthetic video carrying the official short audio losslessly was extracted, transcribed, converted to VTT and round-tripped to SRT. PCM, article outputs, text and times match; pip check passed.
• Two builds identical: 204 HTML pages, 271 files. Changes are confined to two articles, category/archive entries and search data.
• Independent review found no P1/P2. Final changes after review only normalize EOF whitespace and corresponding source hashes; all gates rerun. Signed/DCO and verified rollback bundle saved before push.

The earlier serial full-site run hit a 220-second harness limit and remains recorded as incomplete. Seven article checks and one browser discovery case failed before editing. This is not an accuracy/speed benchmark, real-user-video acceptance, a speaker-identity feature or a PyPI release.