Project to nested columns in StructType::project

[OussamaSaoudi-db]

The current implementation of project and project_as_struct prune a schema down to a subset of its columns. However, they do not prune columns beyond the first level. Consider a checkpoint:

┌──────────────────────┬──────────────────────┬──────────────────────┬──────────────────────┬────────────────────────────────────┐
│         txn          │         add          │        remove        │       metaData       │              protocol              │
│ struct(appid varch…  │ struct(path varcha…  │ struct(path varcha…  │ struct(id varchar,…  │ struct(minreaderversion integer,…  │
├──────────────────────┼──────────────────────┼──────────────────────┼──────────────────────┼────────────────────────────────────┤
│                      │                      │ {'path': part-0000…  │                      │                                    │
│                      │ {'path': part-0000…  │                      │                      │                                    │
│                      │                      │                      │                      │ {'minReaderVersion': 1, 'minWrit…  │
│                      │                      │                      │ {'id': 84b09beb-32…  │                                    │
└──────────────────────┴──────────────────────┴──────────────────────┴──────────────────────┴────────────────────────────────────┘

The current projection would only allow you to project onto txn, add, remove, metaData, or protocol columns. We cannot project onto a nested column like metaData.id. So if we wanted to inspect or filter on a nested column like metaData.id, we could only project it to metaData. In this case, we'd be forced to fetch all the leaf columns of metaData. This is because parquet flattens logical types and only stores the leaf columns.