Rewrite Hyperaudio JSON export as flat words/paragraphs/sections

[maboa]

Summary

Replaces the nested article.section.paragraphs[].spans[] export shape with three flat arrays that are easier for downstream consumers to read, slice, and stream:

{
  "words": [
    { "start": 34.76, "end": 35.4, "text": "Test" }
  ],
  "paragraphs": [
    { "speaker": "SPEAKER_S1", "start": 34.76, "end": 44.68 }
  ],
  "sections": [
    { "start": 0, "end": 239.02, "mediaUrl": "https://example.com/audio.mp3" }
  ]
}

• Times in seconds, each entry carries start and end.
• Speakers live on paragraphs (turns), not on every word — recoverable by checking which paragraph contains a word's start.
• sections[] captures the source-media context. mediaUrl was previously a top-level jsonData.url field; it's now per-section.

Implementation notes

• htmlToJson walks <section> → <p> → <span data-m>, splits speaker spans (class="speaker") out into paragraph.speaker with brackets stripped, and inherits speaker across paragraphs that have no explicit tag (matching the editor's convention).
• jsonToHtml rebuilds the editor HTML using a two-pointer walk to bucket paragraphs into sections and words into paragraphs (sorted by start, handles boundary equality cleanly). Speaker tags are emitted only on speaker change.
• Legacy exports (jsonData.article present) still import via a legacyJsonToHtml fallback so older saved files keep working.
• exportJson() no longer writes the top-level jsonData.url. importJson() reads from sections[0].mediaUrl, falling back to legacy jsonData.url if present.

Closes #296.

Test plan

• Load the default transcript, click Export Hyperaudio JSON → confirm the file contains three flat arrays with the expected shape.
• Import Hyperaudio JSON that file → transcript renders identically (words, speaker labels, paragraph breaks all preserved).
• Edit a word, re-export, re-import → edits round-trip.
• Import a legacy JSON export (if available) → still imports via the fallback.
• Multi-paragraph transcript with multiple speaker turns → speaker labels appear at each turn boundary on import, inherited correctly within a turn.