IterableDataset: split by node and map may preprocess samples that will be skipped anyway

There are two ways an iterable dataset can be split by node:
1. if the number of shards is a factor of number of GPUs: in that case the shards are evenly distributed per GPU
2. otherwise, each GPU iterate on the data and at the end keeps 1 sample out of n(GPUs) - skipping the others.

In case 2. it's therefore possible to have the same examples passed to `prepare_dataset` for each GPU.

This doesn't sound optimized though, because it runs the preprocessing on samples that won't be used in the end.

Could you open a new issue so that we can discuss about this and find a solution ?

_Originally posted by @lhoestq in https://github.com/huggingface/datasets/issues/5360#issuecomment-1592729051_