Inconsistent Arrow Schema When Projecting Nested Parquet File

[tustvold]

Describe the bug

#[test]
  fn test_read_structs() {
      let testdata = arrow::util::test_util::parquet_test_data();
      let path = format!("{}/nested_structs.rust.parquet", testdata);
      let parquet_file_reader =
          SerializedFileReader::try_from(File::open(&path).unwrap()).unwrap();
      let mut arrow_reader = ParquetFileArrowReader::new(Arc::new(parquet_file_reader));
     
      let mut projected = arrow_reader.get_record_reader_by_columns(vec![0], 60).unwrap();

      let batch = projected.next().unwrap().unwrap();
      assert_eq!(projected.schema, batch.schema());

      let schema = batch.schema();
      assert_eq!(batch.column(0).data_type(), schema.field(0).data_type());
  }

This test fails because the schema is computed based on filtering the root nodes.

i.e. It think this is the projected schema

required group field_id=-1 schema {
  required group field_id=-1 roll_num {
    required int64 field_id=-1 min (Int(bitWidth=64, isSigned=true));
    required int64 field_id=-1 max (Int(bitWidth=64, isSigned=true));
    required int64 field_id=-1 mean (Int(bitWidth=64, isSigned=true));
    required int64 field_id=-1 count (Int(bitWidth=64, isSigned=false));
    required int64 field_id=-1 sum (Int(bitWidth=64, isSigned=true));
    required int64 field_id=-1 variance (Int(bitWidth=64, isSigned=true));
  }
}

When it is actually

required group field_id=-1 schema {
  required group field_id=-1 roll_num {
    required int64 field_id=-1 min (Int(bitWidth=64, isSigned=true))
  }
}

This is because the column indices provided to get_record_reader_by_columns is the parquet column indices, not the arrow field index.

To Reproduce

Run test, it fails

Expected behavior

The computed schema should be correct

Additional context

Related to apache/datafusion#2439 and #1651