Unlocking the Potential of Federated Learning: The Symphony of Dataset Distillation via Deep Generative Latents