IterableDataset columns and feature types

Right now, an IterableDataset (e.g. when streaming a dataset) doesn't require to know the list of columns it contains, nor their types: `my_iterable_dataset.features` may be `None`

However it's often interesting to know the column types and types. This helps knowing what's inside your dataset without having to manually check a few examples, and this is useful to prepare a processing pipeline or to train models.

Here are a few cases that lead to `features` being `None`:

1. when loading a dataset with `load_dataset` on CSV, JSON Lines, etc. files: type inference is only done when iterating over the dataset
2. when calling `map`, because we don't know in advance what's the output of the user's function passed to `map`
3. when calling `rename_columns`, `remove_columns`, etc. because they rely on `map`

Things we can consider, for each point above:

1.a infer the type automatically from the first samples on the dataset using prefetching, when the dataset builder doesn't provide the `features`
2.a allow the user to specify the `features` as an argument to `map` (this would be consistent with the non-streaming API)
2.b prefetch the first output value to infer the type
3.a don't rely on `map` directly and reuse the previous `features` and rename/remove the corresponding ones

The thing is that prefetching can take a few seconds, while the operations above are instantaneous since no data are downloaded. Therefore I'm not sure whether this solution may be worth it. Maybe prefetching could also be done when explicitly asked by the user

cc @mariosasko @albertvillanova