fix: keep zero-width chars in table cells when origin is inside

[fahmiridho07]

Fixes #5185.

With quad corrections unset (e.g. after importing pymupdf4llm, which calls TOOLS.unset_quad_corrections(True) at import), a glyph expanded into two characters gets a zero-width bbox for the second one (e.g. the i of fi: (151.70, 56.35, 151.70, 63.35)).

extract_cells keeps a char only if abs(bbox & cell) > 0.5 * abs(bbox). For a zero-area bbox that is 0 > 0, so the char is dropped and table cells lose letters (fine-tuning -> fne-tuning, figures -> fgures), while prose on the same page is fine. Same for zero-width combining overlays (i + U+0338 -> i = j loses the slash).

This change falls back to testing the char origin for empty bboxes:

• empty bbox + origin inside cell -> keep
• empty bbox + origin outside cell -> drop (unchanged behavior for truly outside chars)
• non-empty bbox -> unchanged 50%-overlap test

Verified locally with synthetic zero-width bboxes (pymupdf 1.28.0): old logic drops, new logic keeps inside-origin and still drops outside-origin, no change for normal chars.

[github-actions]

Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.

---

I have read the CLA Document and I hereby sign the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}