Add more OCR models to registry

[davanstrien]

Currently 4 models hardcoded in MODEL_REGISTRY (run.py):

• GLM-OCR (0.9B)
• DeepSeek-OCR (4B)
• LightOnOCR-2 (1B)
• dots.ocr (1.7B)

Candidates to add:

• Qwen2.5-VL-7B — general-purpose VLM with OCR capability
• GOT-OCR2 — dedicated OCR model
• Florence-2 — Microsoft's vision foundation model
• Larger models (7B+) to test whether size helps on hard documents

Also consider making the registry configurable (YAML/TOML file or CLI flag to point at a custom script URL) so users can add their own models without forking.