Export Parquet Tablet Audio-Set is null bytes in Arrow

### Describe the bug

Exporting the processed audio inside the table with the dataset.to_parquet function, the object pyarrow {bytes: null, path: "Some/Path"}
At the same time, the same dataset uploaded to the hub has bit arrays


### Steps to reproduce the bug

1.Get dataset from audio and cast it
2.Export and push dataset
3.It’s scary to be indignant at the difference in the uploaded dataset and the fact that it was saved locally
```py
from datasets import Dataset, Audio
df = Dataset.from_csv("./datasets.csv")
df = df.cast_column("audio", Audio(16000))
df.to_parquet("./datasets.parquet")
df.push_to_hub(repo_id="************", token="**********************")
```

You can use "try replicate case" for this
[replicate_packet.zip](https://github.com/huggingface/datasets/files/15457114/replicate_packet.zip)

### Expected behavior

Two parquet tables identical in content. It is obvious?

### Environment info

Python 3.11+ (I try did it in 3.12 and got same result )