Store DataLoader state on the node
Currently we do not store any state on the nodes. This has an impact on the behaviour of the data loaders, especially if !158 (closed) gets merged.
In particular, if an epoch is not fully exhausted during one round, we lose the information about which data had already been presented to the training, and during the next round we start again from the beginning.
We should have something similar to the Index Generator by Substra.