Render PDF pages to images for vision extraction

[f-wunderlich]

Summary

• render PDF pages to PNG before sending them to vision-capable providers
• guard both the primary and retry vision paths against prompt-only fallbacks when rendering produces no pages
• add regression coverage for the new PDF vision path and retry guard

Why

The existing vision flow can send only the textual prompt for PDFs, which means the model receives no document content.

Testing

• added coverage in tests/test_extraction_modes.py
• added coverage in tests/test_vision_retry.py