Remove unsupported mm_processor_kwargs parameter

The mm_processor_kwargs parameter caused vLLM to fail with OSError
when loading the image processor. Keeping other fixes:
- max_tokens=4096 (recommended by model card)
- enable_prefix_caching=False (recommended by model card)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

lighton-ocr2.py

@@ -341,7 +341,6 @@ def main(
         limit_mm_per_prompt={"image": 1},  # One image per prompt
         enforce_eager=False,  # Use torch.compile for better performance
         enable_prefix_caching=False,  # Recommended by model card
-        mm_processor_kwargs={"cache_gb": 0},  # Disable multimodal processor caching
     )
 
     # LightOnOCR-2 recommended sampling parameters