[Help Wanted] Add ONNX Runtime inference path with PyTorch fallback