A data science team is setting up a new deep learning training environment on NVIDIA GPUs. They already have the CUDA Toolkit installed, but their PyTorch models are running slower than expected during convolution-heavy operations. A colleague suggests ensuring a specific NVIDIA library is properly installed and used by the framework. Which library is specifically designed to provide highly optimized, GPU-accelerated primitives for deep neural network operations such as convolutions, pooling, and activation functions?