Shar: tracking epochs in shard iterator with option for shard re-shuffling each epoch

[pzelasko]

This change is to implement Dan's idea about infinite iteration with re-shuffling of shards at the start of each epoch. I will write a more detailed example/tutorial some time later, but the basic idea is the following snippet:

cuts = (
    CutSet.from_shar(in_dir=shar_dir, shuffle_shards=True, stateful_shuffle=True)
    .repeat()
    .shuffle(buffer_size=10000)
)

yields an infinite CutSet that has two levels of shuffling:

• shard-level: the shards are re-shuffled every time it exhausts the underlying finite CutSet
• cut-level: the cuts are shuffled across shards using a buffer

The rest of the dataloading workflow is identical as with WebDataset. The DataLoader will never stop yielding data so epochs need to be tracked differently. To make it possible, cuts iterated this way have an attached custom field called shar_epoch. In practice when the epoch is incremented, you'd still keep seeing cuts from the previous epoch for a number of steps until they are completely flushed out of the shuffling buffer. If that's undesirable, call .shuffle() first and then .repeat(), but it'll be a bit less random.