Consider adding sharding to the Parquet exported BigQuery content

[fedorov]

As suggested by @mhalle, this can make query from the bucket quicker, perhaps in a significant number of cases. We could shard by collection and perhaps modality.

Related discussion with Claude: https://claude.ai/share/88b80074-de62-4553-a02b-d22d331cf5d2

[mhalle]

Starting with one collection as a test might make sense for performance testing. DuckDB et. al. use the hive path component (collection=nlst) as a column, so you want to drop it from the actual table since the DB will insert it.