fix(detect): Skip dark images drawn behind visible text

[mlissner]

Summary

• Fix false positives from dark images in the PDF structure that are drawn behind other elements (form backgrounds, template layers)
• After confirming the raw image is dark and unicolor, now also renders the page at that location — if the rendered result isn't dark, the image is behind other elements and not hiding anything
• Caught on a real-world bankruptcy form where dark images appeared under the "Revised 05/01/2019" and "Page N" footer text

Test plan

• New test test_image_behind_text_no_results with real-world PDF page
• Verified true positives still detected (image_redaction.pdf, toc_leak_real.pdf)
• All 38 tests pass
• mypy and pre-commit pass
• Merge after feat(detect): Detect TOC/bookmark entries that leak redacted content #212 lands

🤖 Generated with Claude Code

[mlissner]

Folded into #211 — the page-render approach and image_behind_text test are now in that PR directly.