Load weights on CPU before moving to the device (ZeroGPU-compatible)

Add tiny-shakespeare-chat model: weights, model.py, Inference Endpoint handler, card