Model loading fix for vison model

Fix: load base model code (.chat() method) + fine-tuned weights separately

Fix: add transformers+accelerate deps; use float32 on CPU for MiniCPM-V

Fix: use pre-built llama-cpp-python CPU wheel to avoid OOM during build

Fix: pydantic>=2.11.10 for gradio MCP extra compatibility

Fix: bump huggingface_hub to >=0.33.5 for gradio 6.16.0 compatibility

All models training uploaded

Finetuning completed for yolo26n-indian-fmcg-detection and minicpm5-1b-indian-fmcg-normalizer