PyTorch native quantization and sparsity for training and inference