Use embedded PDF bookmarks in Flash trees

[rejojer]

PageIndex Flash now reads the PDF's embedded bookmarks (outline) and merges them into the extracted tree. No LLM involved; classification and merging are rule-based.

• Trustworthy deep bookmark sets become the tree frame, and detected sections they lack are grafted back in by page range after noise filtering.
• Coarse (chapter-level) sets become the chapter frame: detected nodes are re-hung under them, garbled extracted titles are repaired from the bookmark strings, and missing entries are filled in when the page text confirms them.
• Junk sets (generic "Page 12" titles, file-name enumerations, non-monotonic targets) are ignored and detection stands alone.
• On by default for page_index_flash / extract_toc; disable with use_embedded_toc=False or --no-embedded-toc. The result gains a toc_source key ("bookmarks", "hybrid", or "detected").
• Bookmarks can also structure documents whose text layer is too poor for detection.

Title matching normalizes numbering notation (roman and arabic compare equal) and handles CJK titles.