[Proposal] Evaluate Arrow IPC format for reader

[wolfgang-desalvador]

This issue tracks the proposal of evaluating Apache Arrow IPC as an alternative on‑disk data representation format for the MLPerf Storage reader pipeline, with the goal of improving effective data loading reducing CPU overhead during sample deserialization.

[wolfgang-desalvador]

Relates to mlcommons/DLIO_local_changes#14

[wolfgang-desalvador]

Analysis -> https://gist.github.com/wolfgang-desalvador/cc0679dc772fe9eeb3b1b69214015eb2

[russfellows]

I am continue to be inclined to prefer Arrow IPC over parquet, for two reasons:

1. Wolfgang's research is solid, and indicates real issues with Parquet for high throughput data loading, my research concurs
2. Real practitioners (aka @bbelgodere ), confirmed they moved away from Parquet and towards Arrow IPC for these reasons.