Boosting PyTorch Inference on CPU: From Post-Training Quantization to Multithreading