For data-only datasets, streaming and non-streaming don't behave the same

See https://huggingface.co/datasets/huggingface/transformers-metadata: it only contains two JSON files.

In streaming mode, the files are concatenated, and thus the rows might be dictionaries with different keys:

```python
import datasets as ds
iterable_dataset = ds.load_dataset("huggingface/transformers-metadata", split="train", streaming=True);
rows = list(iterable_dataset.take(100))
rows[0]
# {'model_type': 'albert', 'pytorch': True, 'tensorflow': True, 'flax': True, 'processor': 'AutoTokenizer'}
rows[99]
# {'model_class': 'BartModel', 'pipeline_tag': 'feature-extraction', 'auto_class': 'AutoModel'}
```

In normal mode, an exception is thrown:

```python
import datasets as ds
dataset = ds.load_dataset("huggingface/transformers-metadata", split="train");
```

```
ValueError: Couldn't cast
model_class: string
pipeline_tag: string
auto_class: string
to
{'model_type': Value(dtype='string', id=None), 'pytorch': Value(dtype='bool', id=None), 'tensorflow': Value(dtype='bool', id=None), 'flax': Value(dtype='bool', id=None), 'processor': Value(dtype='string', id=None)}
because column names don't match
```