[GCS] Scanning hive-partitioned dataset results in orders of magnitude more network traffic than it should

### Describe the bug, including details regarding any error messages, version, and platform.

I have a hive-partitioned dataset in a Google Cloud Storage bucket. Its size is around 54MB according to `gsutil du` and verified by downloading it and checking locally. However, if I open it with `ds = pyarrow.dataset.dataset("gs://...", partitioning="hive", format="parquet")` and then traverse it with `ds.to_batches()`, it results in multiple GB of inbound network traffic and takes much longer than simply downloading the data.

This is with PyArrow 10.0.1. MacOS 12.6 with Apple M1 CPU.

### Component(s)

Python

Reproduce the report with a hive-partitioned Parquet dataset in Google Cloud Storage using pyarrow.dataset.dataset(...), then traverse it with ds.to_batches(). Compare the observed network traffic and runtime with the dataset's roughly 54 MB size and a direct download; done means scanning no longer causes multi-GB traffic.

Written by the indexing model from the issue text.