[codex] Refactor AKY24IO PRF batching

[SoraSuegami]

Summary

• Replace AKY24IO public PRF seed-bit configuration with PRF batch bits.
• Derive PRF round counts from input size divided by batch bits, and derive branch counts as 2^batch_bits.
• Align benchmark cascade stages, simulation seed sizing, and GPU bench configuration with batched PRF rounds.
• Reuse representative benchmark units and clarify noise-refresh combine comments.

Motivation

AKY24IO previously treated PRF seed-refresh rounds as an independent public seed-bit count. That made simulation and benchmark estimation inconsistent with the batched PRF model used by DiamondIO-style IO flows.

Validation

• cargo +nightly fmt --all
• FIDESLIB_SKIP_SUBMODULE_UPDATE=1 cargo test --lib aky24_io