Describe the bug

When persistent workers are enabled, the epoch that's set via the IterableDataset instance held by the training process is ignored by the workers as they are disconnected across processes.

PyTorch samplers for non-iterable datasets have a mechanism to sync this, datasets.IterableDataset does not.

In my own use of IterableDatasets I usually track the epoch count which crosses process boundaries in a multiprocessing.Value

Steps to reproduce the bug

Use a streaming dataset (Iterable) w/ the recommended pattern below and persistent_workers=True in the torch DataLoader.

for epoch in range(epochs):
    shuffled_dataset.set_epoch(epoch)
    for example in shuffled_dataset:
        ...

Expected behavior

When the canonical bit of code above is used with num_workers > 0 and persistent_workers=True, the epoch set via set_epoch() is propagated to the IterableDataset instances in the worker processes

Environment info

N/A