Benchmark randomness vs performance tradeoff

[ilan-gold]

It is clear (see https://arxiv.org/pdf/2506.01883) that there is a tradeoff in block-based loaders between randomness and read throughput. Generally, more randomness entails less read throughput.

• At what chunk size on-disk does either pre-shuffling become unnecessary? Is this chunk size performant and what is the tradeoff?