ARQ: A Mixed-Precision Quantization Framework for Accurate and Certifiably Robust DNNs