Prune nested leaves in the projection fallback for files without field ids #3480

Is your feature request related to a problem or challenge?

For a Parquet file with no field ids and no name mapping, get_arrow_projection_mask_fallback projects each requested top-level column whole, so the reader decodes every leaf under it.

After #3437 (fix for #3436), fields inside a struct of such a file have no id and read as null, as in iceberg-java. Their leaves are still decoded and then discarded by RecordBatchTransformer. For wide or deeply nested struct columns this is wasted I/O and decoding.

Raised in review of #3437.

Describe the solution you'd like

Project only the leaves the result needs:

• Primitive top-level columns, and the element of a list and the key and value of a map, which match by position and keep their values.
• One leaf per struct. A struct keeps the file's validity, which Parquet stores in the definition levels of its leaves, so at least one leaf must still be read. Picking a cheap leaf (for example the narrowest primitive) would keep the cost low.

The read result should not change; only the projection mask does.

Willingness to contribute

I would be willing to contribute to this feature with guidance from the Iceberg community