Replace PyMuPDF (AGPL-3.0) with pdf2image (MIT) — preserves vision pipeline

[BrandonS7]

What this does

PyMuPDF is dual-licensed under AGPL-3.0 or Artifex Commercial. The project is MIT-licensed. When AGPL code is used in a network-accessible application, it requires the entire application source to be made available under AGPL. This creates a license incompatibility.

Replaced PyMuPDF with pdf2image (MIT) + Pillow (PIL License) for PDF page rendering:

• pdf2image renders PDF pages to PIL Images using poppler (same quality)
• Pillow handles PNG conversion for base64 encoding
• The GPT-4 Vision analysis pipeline is fully preserved — pages are still rendered as images and sent to Vision API for content extraction
• CLI interface, streaming interface, DPI parameter, and embedding generation all unchanged

Dependencies

• Removed: PyMuPDF==1.26.6 (AGPL-3.0)
• Added: pdf2image>=1.17.0 (MIT), Pillow>=10.0.0 (PIL/MIT-compatible)
• Note: pdf2image requires poppler-utils on the system (typically available in Alpine/Debian Docker images via apk add poppler-utils or apt install poppler-utils)

What did NOT change

• Vision model and prompt
• OpenAI API integration
• Embedding generation
• File upload handling
• Any other backend code