feat(zerogpu): clean CPU-to-CUDA model movement and teardown in @spaces.GPU

fix(scores): parse score criteria into list of levels

fix(vram): enforce single model in VRAM and flexible criteria parsing

fix: order 3b model detection before 7b fallback in resolve_model_id

feat: use verified Qwen2.5-Coder-3B and Qwen2.5-7B-abliterated models

fix: make chat_completions synchronous and return traceback on error

fix: deploy 3-model hybrid engine with ZeroGPU startup handshake

fix: parse ChatInterface messages and launch uvicorn for root FastAPI endpoints

fix: import spaces before torch and initialize ZeroGPU with demo.launch

fix: decorate Gradio handlers with @spaces.GPU and remove pypi spaces from requirements

feat: deploy Laya System 1 + 3B abliterated + 7B abliterated super-stack