cuDNN Toggle (AMD-aware)

Toggle torch.backends.cudnn.enabled ON or OFF, but only applies the change when an AMD-like environment is detected (AMD/ZLUDA/HIP). On NVIDIA (or when AMD is not detected), this node will not modify cuDNN and will print a notice. If you care about preserving the original state of cuDNN, connect the prev_cudnn output of the disabling node to the enable_cudnn input of the enabling node. To ensure correct behavior, connect the any_input input and any_output output of the node to intercept an IMAGE or LATENT immediately before and after the node you wish to affect. The most common use case is to disable cuDNN for VAE Decode and VAE Encode nodes, which can receive significant speedups on AMD-like setups.