Recover truncated inline-image PDF text per page with optional PyMuPDF

[Çağdaş Yürekli (cagdasyurekli)]

PDF text can stop at an inline image while conversion still returns a plausible prefix. This adds optional, local, per-page recovery through markitdown[pdf,pdf-recovery].

The recovery backend inspects decoded images, including compressed page streams, and replaces a plain page only when its extracted words strictly extend the primary word sequence. Empty primary pages are eligible. Pages recognized as forms/tables retain their existing Markdown, and there is no whole-document length gate or replacement. Backend absence/failure keeps the primary extraction. PyMuPDF is not added to the all or pdf extras.

Relates to #1870; alternative to #1883, #1889 and #2092, incorporating the per-page preservation and compressed-stream review concerns. This is not a claim that every affected PDF is recovered: reordered text and form/table pages are deliberately left alone, and there is no OCR.

Validation:

• Three real-PDF regressions fail on main and pass here: uncompressed inline content, compressed inline content, and a mixed document with over 2,048 characters of healthy prose plus a Markdown table.
• The synthetic fixture uses an ASCII85 inline image with a malformed bare ~ terminator that PyMuPDF tolerates while pdfminer consumes following text. Fixture provenance is included. This demonstrates the parser-divergence case, not that all valid inline images cause truncation.
• Nine focused tests pass, including empty primary text, rejecting unrelated longer output, optional-dependency absence, and backend failure.
• Full core suite: 975 passed, 43 skipped on Python 3.12 with PyMuPDF 1.28.2. Remote/API tests skipped. Black 23.7.0 and git diff --check pass.

Trade-off: installing the optional backend adds a second local PDF inspection pass. Without it, the normal extraction path is retained.

[Çağdaş Yürekli (cagdasyurekli)]

Updated the branch with current main (4cc9fa1) using a locally verified merge.

Validation on Python 3.12 with PyMuPDF 1.28.2: Black 23.7.0 and git diff --check pass; full core pytest: 1059 passed, 57 skipped under the repository CI skip policy. The optional PDF recovery regression tests ran successfully.