Row groups are read out of order or with completely different values

### Describe the bug

Datafusion is reading row groups out of order and sometimes with completely different values for the row groups. The data is verified by reading the same files using the Python `pyarrow.parquet` library.

The `pyarrrow` and `datafusion` reader read the same values when the file has 126 row groups. But give complete different values when the file has 127 row groups.

### To Reproduce

https://github.com/nautechsystems/nautilus_experiments/tree/datafusion-bug

The steps, data and results are documented in this repo and branch. The README is shared here again.

---

Use the python script to extract row group information from the parquet files using pyarrow.

```bash
pip install -r requirements.txt
python extract_ts_init.py 126-groups.parquet 126-groups-python.csv
python extract_ts_init.py 127-groups.parquet 127-groups-python.csv
```

Run the rust executable to extract row group information from the parquet files using datafusion.

```bash
cargo run 126-groups.parquet > 126-groups-rust.csv
cargo run 127-groups.parquet > 127-groups-rust.csv
```

Ideally there should be no difference between the csv files for the row groups. However, 126 works properly. But 127 gives different results for Python and Rust.

This shows that indeed there's no difference with 126 groups.

```bash
diff 126-groups-rust.csv 126-groups-python.csv # no diff
diff 126-groups-rust.csv 126-groups-python.csv # big diff, things crazy
```

We can also make sure that these are in fact from the same data source with just one extra row group with this command which shows 127 groups python has only one extra entry at the end.

```bash
diff 126-groups-python.csv 127-groups-python.csv
```

### Expected behavior

Datafusion reader should
* read row groups in order they were written
* read the same values for the same row group even when the file increases in size
* read the same values as the python `pyarrow` parquet reader

### Additional context

_No response_

Start with the reproduction repository and branch linked in the issue, then run extract_ts_init.py on the 126-groups.parquet and 127-groups.parquet files and compare its CSV output with the Rust cargo run output. Trace the DataFusion reader path used by that executable; done means matching row-group order and values across both files and the Python results.

Written by the indexing model from the issue text.