Align IterableDataset.shuffle with Dataset.shuffle

[lhoestq]

From #3444 , Dataset.shuffle can have the same API than IterableDataset.shuffle (i.e. in streaming mode).

Currently you can pass an optional seed to both if you want, BUT currently IterableDataset.shuffle always requires a buffer_size, used for approximate shuffling. I propose using a reasonable default value (maybe 1000) instead.

In this PR, I set the default buffer_size value to 1,000, and I reorder the IterableDataset.shuffle arguments to match Dataset.shuffle, i.e. making seed the first argument.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[mariosasko]

We should also add generator as a param to shuffle to fully align the APIs, no?

[lhoestq]

I added the generator argument.

I had to make a few other adjustments to make it work. In particular when you call set_epoch() on a streaming dataset, it updates the underlying random generator by using a new effective seed. The effective seed is generated using the previous generator and the epoch number.

[mariosasko]

Thanks! One comment: