Fix context-aware PDF annotation disambiguation

[tenglvjun]

Summary

• preserve bounded visible context around Markdown selections and use it with page hints to select the correct occurrence in the PDF
• resolve existing Zotero PDF annotations to a unique Markdown range using pageIndex, the source offset encoded in sortIndex, and up to 80 Unicode code points of PDF context
• apply context disambiguation to both exact and normalized text matches while preserving ambiguous results when context or the local PDF index is unavailable
• improve PDF/Markdown normalization for LaTeX relational operators, temperature units, and related extraction differences
• document the bidirectional matching behavior in English and Simplified Chinese

Verification

• npm run check
• npm test (1156/1156 passed)
• npm run build
• replayed Zotero annotation 905 against the reported wrist-temperature paper: one match at Markdown range 1096..1270, zero unmatched annotations
• verified the generated XPI contains the runtime text-quote wiring
• kept all version fields at 0.3.0