fix: OCR scanned pages when mixed with text pages

[Brumbelow]

Fixes #1791

• When a multi-page PDF mixed selectable text with scanned pages, the converter only emitted content for pages that had either an image object or an extractable text layer. A scanned page with neither (no page.images, no page.extract_text()) silently fell through, and the whole-document _ocr_full_pages fallback only runs when every page is empty. So pages 2..N just disappeared.

• This adds a small _ocr_page helper that renders a single pdfplumber page to PNG and runs the configured OCR service on it, and calls it from the per-page loop when the page yields nothing the existing branches can use. The output format matches what _ocr_full_pages already emits, so existing snapshot tests are unaffected.

Test: new pdf_text_then_blank.pdf fixture (~1.7 KB, page 1 has selectable text, page 2 is blank) plus a snapshot test that asserts both pages survive the round-trip.

[Brumbelow]

@microsoft-github-policy-service agree