Please describe why this is necessary.

Delta features such as deletion vectors and row tracking rely on file-local row index information. Today, we implicitly rely on the row positions of returned data to match those rows' positions in the parquet file they were read from. That assumption precludes optimizations such as row group skipping that would break the relationship between row ordinals in the returned data and row indexes in the underlying file.

Describe the functionality you are proposing.

Add row index support to the default parquet reader as a first step to propagating row indexes through the kernel (deletion vectors, in particular, since they're already implemented).

Additional context

No response