Diversity-Enforced Latent Representations in AutoEncoder Ensembles for Cross-Dataset Network Intrusion Detection