Fix: pin gradio==5.12.0 to avoid API schema bug

Switch to transformers CPU inference (Qwen3.5 not yet supported by llama.cpp)

Initial Space: Gradio demo with GGUF CPU inference

initial commit