QONNX: Representing Arbitrary-Precision Quantized Neural Networks