None batched `with_transform`, `set_transform`

**Is your feature request related to a problem? Please describe.**

A `torch.utils.data.Dataset.__getitem__` operates on a single example.
But 🤗 `Datasets.with_transform` doesn't seem to allow non-batched transform.

**Describe the solution you'd like**

Have a `batched=True` argument in `Datasets.with_transform`

**Describe alternatives you've considered**

* Convert a non-batched transform function to batched one myself.
* Wrap a 🤗 Dataset with torch Dataset, and add a `__getitem__`. 🙄
* Have `lazy=False` in `Dataset.map`, and returns a `LazyDataset` if `lazy=True`. This way the same `map` interface can be used, and existing code can be updated with one argument change.