np arrays get turned into native lists

## Describe the bug
When attaching an `np.array` field, it seems that it automatically gets turned into a list (see below). Why is this happening? Could it lose precision? Is there a way to make sure this doesn't happen?

## Steps to reproduce the bug
```python
>>> import datasets, numpy as np
>>> dataset = datasets.load_dataset("glue", "mrpc")["validation"]
Reusing dataset glue (...)
100%|███████████████████████████████████████████████| 3/3 [00:00<00:00, 1360.61it/s]
>>> dataset2 = dataset.map(lambda x: {"tmp": np.array([0.5])}, batched=False)
100%|██████████████████████████████████████████| 408/408 [00:00<00:00, 10819.97ex/s]
>>> dataset2[0]["tmp"]
[0.5]
>>> type(dataset2[0]["tmp"])

```

## Expected results
`dataset2[0]["tmp"]` should be an `np.ndarray`.

## Actual results
It's a list.

## Environment info

- `datasets` version: 2.3.2
- Platform: mac, though I'm pretty sure it happens on a linux machine too
- Python version: 3.9.7
- PyArrow version: 6.0.1