[Feature Request] Multi resolution training scheduling

[StableLlama]

Adding my training data in the multidata backend in high and low resolution it's already easy to give the low-res a higher repeat to show it more often to the trainer and thus speed the training up.

But it would be great, i.e. the feature request, to also set an epoch number there. Then this data entry would be ignored up to that epoch and only be used during later epochs.

The result would be that I'm starting with only the low res images up to an epoch and then from that point on the low and the high res is used.

Additionally and most likely experimentally it might be interesting to limit a data set (like the high-res one) to the later steps during training as during the early steps the high-res information would be hidden in the noise anyway.