Projected BNNs: Avoiding Weight-space Pathologies by Learning Latent Representations of Neural Network Weights