Improving Post Training Neural Quantization: Layer-wise Calibration and Integer Programming