Preserve row counts for zero-column tables in IO readers

Follow-up to the zero-column support #22765.

The core PR made zero-column tables (shape `(N, 0)`) representable and threaded their row counts through table-producing algorithms, interop, and `plc.Table`. The remaining gap is in the IO readers, which still collapse zero-column reads to `(0, 0)` and lose the row count.

## To do

* [ ] Parquet reader (including chunked and filtered read paths)
* [ ] ORC reader (including chunked read paths)
* [ ] CSV reader
* [ ] AVRO, JSON, and experimental Parquet paths (hybrid scan, deletion vectors) where applicable
* [ ] Unskip the placeholder reader tests (`ParquetReaderTest`, `OrcReaderTest`, and `CsvReaderTest.ZeroColumnsPreservesRowCount`)
* [ ] Remove the cudf-polars workarounds: `DataFrame._num_rows_override`, the `DataFrameScan` zero-width path, `Scan._get_parquet_row_count_from_metadata`, and the related xfails in `inject_gpu_engine.py`

Start by locating the ParquetReaderTest, OrcReaderTest, and CsvReaderTest.ZeroColumnsPreservesRowCount placeholders, then trace the Parquet, ORC, CSV, AVRO, JSON, and experimental Parquet reader paths listed in the issue. Done means zero-column reads preserve their row counts across the applicable chunked and filtered paths, all listed tests are unskipped, and the named cudf-polars workarounds and related xfails are removed.

Written by the indexing model from the issue text.