Investigate why Rust outperforms Zig by 40% on 32B inputs

On most input sizes, Zig leads by ~10%, but on 32B inputs Rust is unexpectedly 40% faster. This anomaly needs investigation.

See benchmark results: [results/2025-09-13-23-49-36.md](https://github.com/jadnohra/zig-crypto-bench/blob/main/results/2025-09-13-23-49-36.md#sha256-32-b)

Results showing the anomaly:
- Zig stdlib: 42 ns/op (726.61 MB/s)
- Rust (sha2): 29 ns/op (1052.33 MB/s) - **1.45x faster**

Possible causes to investigate:
- FFI overhead differences (Rust called via FFI, Zig runs natively)
- Timer measurement noise at ~25-29ns durations
- Different startup/initialization overhead
- Branch prediction patterns
- Cache alignment effects

Need to determine if this is reproducible across different hardware or just measurement noise.

Start with results/2025-09-13-23-49-36.md and inspect the SHA-256 32B benchmark results. Reproduce the comparison across different hardware while checking the listed possibilities, including FFI overhead and timer noise. Done means determining whether the 40% Rust advantage is reproducible or measurement noise.

Written by the indexing model from the issue text.