Use parallel parquet read with `datafusion>=55`

Current implementation of `npd.read_parquet(..., engine="datafusion")` requires `datafusion.execution.target_partitions == 1` setting so the result row order i stable and corresponds to the order in the file. With DataFusion v55 (released in Rust, but not in Python yet), we will be able to remove this setting, and use `.to_array_table` call with `.collect_partitioned`, which will run in parallel and persist row order if the v55 new setting `datafusion.execution.enable_file_stream_work_stealing` is applied.

Side note: we should refactor `io.py` to a subpackage, see https://github.com/lincc-frameworks/nested-pandas/pull/531/changes#r3962692931

Start at the npd.read_parquet implementation in io.py and trace the datafusion engine path. Review the DataFusion v55 APIs mentioned, especially to_array_table, collect_partitioned, and enable_file_stream_work_stealing; done means parallel reads preserve file row order without requiring target_partitions == 1.

Written by the indexing model from the issue text.