[Python] File reading regression

After ARROW-6626 was merged, we're seeing a slow down in file reading benchmarks:

One example: https://conbench.ursa.dev/benchmarks/b92e91fe8a4041148360d9433552277b/

**Reporter**: [Jonathan Keane](https://issues.apache.org/jira/browse/ARROW-14187) / @jonkeane
#### PRs and other links:
- [GitHub Pull Request #11284](https://github.com/apache/arrow/pull/11284)

_{**Note**: *This issue was originally created as [ARROW-14187](https://issues.apache.org/jira/browse/ARROW-14187). Please see the [migration documentation](https://github.com/apache/arrow/issues/14542) for further details.*}

Start with the linked Conbench benchmark and compare the file-reading results before and after ARROW-6626. Review GitHub pull request #11284 for the investigation and proposed resolution; done means identifying and resolving the regression so the affected benchmark returns to its prior performance.

Written by the indexing model from the issue text.