Feature Request: Support for Pandas `Categorical`

```
from datasets import Dataset
import pandas as pd
import pyarrow

df = pd.DataFrame(pd.Series(["a", "b", "c", "a"], dtype="category"))
pyarrow.Table.from_pandas(df)
Dataset.from_pandas(df)
# Throws NotImplementedError
# TODO(thom) this will need access to the dictionary as well (for labels). I.e. to the py_table
```

I'm curious if https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L796 could be built out in a way similar to `Sequence`?

e.g. a `Map` class (or whatever name the maintainers might prefer) that can accept:

```
index_type = generate_from_arrow_type(pa_type.index_type)
value_type = generate_from_arrow_type(pa_type.value_type)
```

and then additional code points to modify:

- FeatureType: https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L694
- A branch to handle Map in get_nested_type: https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L719
- I don't quite understand what `encode_nested_example` does but perhaps a branch there? https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L755
- Similarly, I don't quite understand why `Sequence` is used this way in `generate_from_dict`, but perhaps a branch here? https://github.com/huggingface/datasets/blob/master/src/datasets/features.py#L775

I couldn't find other usages of `Sequence` outside of defining specific datasets, so I'm not sure if that's a comprehensive set of touchpoints.