Fix ZeroGPU startup and local GPU inference path

Keep Space alive when model weights are unavailable

Implement ZeroGPU Space runtime

align project with CLAUDE spec and hf space deploy