DISC-GAN: Disentangling Style and Content for Cluster-Specific Synthetic Underwater Image Generation