concatenate_datasets removes ClassLabel typing.

[Dref360]

Describe the bug

When concatenating two datasets, we lose typing of ClassLabel columns.

I can work on this if this is a legitimate bug,

Steps to reproduce the bug

import datasets
from datasets import Dataset, ClassLabel, Value, concatenate_datasets

DS_LEN = 100
my_dataset = Dataset.from_dict(
    {
        "sentence": [f"{chr(i % 10)}" for i in range(DS_LEN)],
        "label": [i % 2 for i in range(DS_LEN)]
    }
)
my_predictions = Dataset.from_dict(
    {
        "pred": [(i + 1) % 2 for i in range(DS_LEN)]
    }
)

my_dataset = my_dataset.cast(datasets.Features({"sentence": Value("string"), "label": ClassLabel(2, names=["POS", "NEG"])}))
print("Original")
print(my_dataset)
print(my_dataset.features)


concat_ds = concatenate_datasets([my_dataset, my_predictions], axis=1)
print("Concatenated")
print(concat_ds)
print(concat_ds.features)

Expected results

The features of concat_ds should contain ClassLabel.

Actual results

On master, I get:

{'sentence': Value(dtype='string', id=None), 'label': Value(dtype='int64', id=None), 'pred': Value(dtype='int64', id=None)}

Environment info

• datasets version: 1.14.1.dev0
• Platform: macOS-10.15.7-x86_64-i386-64bit
• Python version: 3.8.11
• PyArrow version: 4.0.1

[Dref360]

Something like this would fix it I think: /~https://github.com/huggingface/datasets/compare/master...Dref360:HF-3111/concatenate_types?expand=1