Gather-pq-parts takes a while

I'm reading a large parquet dataset with 128 MiB files and finding that `gather-pq-parts` takes a while when first creating the dataset. I'm wondering under what conditions this is triggered. I tried looking through the code, but it was fairly complex.

@rjzamora any help?

Start by tracing where `gather-pq-parts` runs during initial creation of a dataset, using the reported 128 MiB parquet files as the reproduction case. Compare dataset creation with and without the conditions that trigger it; done means documenting those conditions and explaining the reported delay.

Written by the indexing model from the issue text.