Formatted map + with_format(None) changes array dtype for iterable datasets

### Describe the bug

When applying with_format -> map -> with_format(None), array dtypes seem to change, even if features are passed

### Steps to reproduce the bug

```python
features=Features(**{"array0": Array3D((None, 10, 10), dtype="float32")})
dataset = Dataset.from_dict({f"array0": [np.zeros((100,10,10), dtype=np.float32)]*25}, features=features)
ds = dataset.to_iterable_dataset().with_format("numpy").map(lambda x: x, features=features)
ex_0 = next(iter(ds))
ds = dataset.to_iterable_dataset().with_format("numpy").map(lambda x: x, features=features).with_format(None)
ex_1 = next(iter(ds))
assert ex_1["array0"].dtype == ex_0["array0"].dtype, f"{ex_1['array0'].dtype} {ex_0['array0'].dtype}"
```

### Expected behavior

Dtypes should be preserved.

### Environment info

3.0.2