Hyper-Representations as Generative Models: Sampling Unseen Neural Network Weights