Possible memory leak with to_arrow_batch_reader()

### Apache Iceberg version

0.9.1 (latest release)

### Please describe the bug 🐞

## Summary
It seems that there is memory leak when running to_arrow_batch_reader(), it takes ~30GB memory to read an iceberg table with single 40MB parquet files

Example code:
```python
import boto3
from pyiceberg.table import StaticTable, Table

def iceberg_table_from_metadata_path(metadata_path: str) -> StaticTable:
session = boto3.Session(region_name="")
credentials = session.get_credentials()
credentials = credentials.get_frozen_credentials()
table = StaticTable.from_metadata(
metadata_path,
{
"client.secret-access-key": credentials.secret_key,
"client.access-key-id": credentials.access_key,
"client.session-token": credentials.token,
"client.region": "",
},
)
return table

def main():
metadata_path = ".metadata.json"
iceberg_table = iceberg_table_from_metadata_path(metadata_path)
scan_kwargs = {"row_filter": f"PARTITION='train'"}
batch_reader = iceberg_table.scan(**scan_kwargs).to_arrow_batch_reader()
for batch in batch_reader:
print(f"Inside batch reader")

plan_files = iceberg_table.scan(**scan_kwargs).plan_files()
for file in plan_files:
print(file.file.file_path)

print("Hello from pyiceberg-test!")

if __name__ == "__main__":
main()

```

Running using memray
```
uv run memray run main.py
```

Charts

Dependencies
```
"boto3>=1.40.21",
"memray>=1.18.0",
"pyarrow>=21.0.0",
"pyiceberg==0.10.0rc1",
"s3fs>=0.4.2"
```

### Willingness to contribute

- [ ] I can contribute a fix for this bug independently
- [x] I would be willing to contribute a fix for this bug with guidance from the Iceberg community
- [ ] I cannot contribute a fix for this bug at this time

Commencez par reproduire l’exemple de main.py avec memray run main.py, en vous concentrant sur le point d’entrée to_arrow_batch_reader() et en comparant la mémoire pendant l’itération par lots avec l’appel ultérieur à plan_files(). Suivez le chemin du scan et du lecteur Arrow pour déterminer si la mémoire reste retenue ; le travail est considéré comme terminé lorsque l’exemple lit la table sans croissance illimitée et que le comportement est couvert par un test de régression approprié.

Rédigé par le modèle d'indexation à partir du texte de l'issue.