Patch: fix dependency/runtime fallbacks for OCR models

Add cyttic OCR models to registry

Use local ZeroGPU offline inference only

Add PaddleOCR + DeepSeek OCR models and raise default output token budget

Fix runtime dependency compatibility and add Nemotron-OCR-v2

Add Qwen3-VL 4B and 30B-A3B thinking/instruct variants

Upload Hebrew OCR comparison Space (printed + handwritten Hebrew OCR benchmark)