Cross-Layer Approximation For Printed Machine Learning Circuits