Expanding streaming capabilities

Some ideas for a few features that could be useful when working with large datasets in streaming mode.

## `filter` for `IterableDataset`
Adding filtering to streaming datasets would be useful in several scenarios:
- filter a dataset with many languages for a subset of languages
- filter a dataset for specific licenses
- other custom logic to get a subset
The only way to achieve this at the moment is I think through writing a custom loading script and implementing filters there.

## `IterableDataset` to `Dataset` conversion
In combination with the above filter a functionality to "play" the whole stream would be useful. The motivation is that often one might filter the dataset to get a manageable size for experimentation. In that case streaming mode is no longer necessary as the filtered dataset is small enough and it would be useful to be able to play through the whole stream to create a normal `Dataset` with all its benefits.

```python
ds = load_dataset("some_large_dataset", streaming=True)
ds_filter = ds.filter(lambda x: x["lang"]="fr")
ds_filter = ds_filter.stream() # here the `IterableDataset` is converted to a `Dataset`
```
Naturally, this could be expanded with `stream(n=1000)` which creates a `Dataset` with the first `n` elements similar to `take`.

## Stream to the Hub
While streaming allows to use a dataset as is without saving the whole dataset on the local machine it is currently not possible to process a dataset and add it to the hub. The only way to do this is by downloading the full dataset and saving the processed dataset again before pushing them to the hub. The API could looks something like:

```python
ds = load_dataset("some_large_dataset", streaming=True)
ds_filter = ds.filter(some_filter_func)
ds_processed = ds_filter.map(some_processing_func)
ds_processed.push_to_hub("new_better_dataset", batch_size=100_000)
```

Under the hood this could be done by processing and aggregating `batch_size` elements and then pushing that batch as a single file to the hub. With this functionality one could process and create TB scale datasets while only requiring size of `batch_size` local disk space.

cc @lhoestq @albertvillanova