Writing parquet fails modest memory limit

### Describe the bug

As discussed on discord, here's another external sort usecase that's failing.

Repro:
https://github.com/ivankelly/df-repro

To run:
```
$ bash setup.sh # download the source data
$ RUST_LOG=trace cargo run
...
Error: Resources exhausted: Failed to allocate additional 1450451 bytes for ParquetSink(ArrowColumnWriter) with 62770337 bytes already allocated for this reservation - 1107184 bytes remain available for the total pool
```

The code reads in a bunch of parquet files (889MB in total) and tries to sort and output to a single parquet file.
Memory is limited to 100MB.
Different batch sizes and target partitions doesn't help.

### To Reproduce

_No response_

### Expected behavior

_No response_

### Additional context

_No response_

Start with the external reproduction repository and run bash setup.sh followed by RUST_LOG=trace cargo run under the 100MB limit. Trace the failure at ParquetSink(ArrowColumnWriter) during the external sort, then verify that the parquet files can be sorted and written to one output file without exhausting the memory reservation.

Written by the indexing model from the issue text.