Make Dataset streaming queries retryable

### Feature request

Streaming datasets, as intended, do not load the entire dataset in memory or disk. However, while querying the next data chunk from the remote, sometimes it is possible that the service is down or there might be other issues that may cause the query to fail. In such a scenario, it would be nice to make these queries retryable (perhaps with a backoff strategy).

### Motivation

I was working on a model and the model checkpoints after every 1000 steps. At step 1800 I got a 504 HTTP status code error from Huggingface hub for my pytorch `dataloader`. Given the size of my model and data, it took around 2 hours to reach 1800 steps and now it will take about an hour to recover the lost 800. It would be better to get a retryable querying strategy.

### Your contribution

It would be better if someone having experience in this area takes this up as this would require some testing.