Overwrite Quantization: Opportunistic Outlier Handling for Neural Network Accelerators