Cache latents

To speed up the LoRA training process, pre-generate the latent representation of the training images in advance, and save into system memory.