Reuse Parquet OffsetIndex objects and direct-plan decisions

### Problem

Selective Parquet reads repeat CPU work within one file reader: Arrow 17 reparses OffsetIndex on each lookup, page planning scans page locations for sparse row selections, and decoding rebuilds a direct-read plan just to determine whether compressed row coordinates apply.

The byte cache in #272 reduces repeated I/O but does not eliminate this parsing and planning work.

### Proposed improvement

- Reuse parsed OffsetIndex objects within each retained row-group reader.
- Seek across page gaps for sparse selections and visit each selected page once.
- Reuse the direct-plan decision during decoding.

Preserve index validation, dictionary handling, per-leaf row coordinates and missing-index fallbacks. Keep parsed objects reader-local without introducing a shared data cache or public option.

### Validation

Check selected values and index lifetimes, including sparse/dense selections and the row-group retention limit. Compare against main with the format benchmark, reporting latency and storage bytes without assuming a general speedup.

Implementation: #314.

Commencez par examiner l'issue #314, qui est désignée comme l'implémentation de ce travail. Validez les valeurs sélectionnées et les durées de vie des index pour des sélections sparse et dense, y compris les limites de rétention des row-groups, puis comparez le format benchmark avec main en termes de latence et d'octets de stockage, sans supposer un gain de vitesse général.

Rédigé par le modèle d'indexation à partir du texte de l'issue.