BLAKE3 cryptographic hash — self-built single-thread implementation. Follows the spec at https://github.com/BLAKE3-team/BLAKE3-specs/blob/master/blake3.pdf.
Scope: unkeyed hash(input) -> [u8; 32] only. KDF / keyed-hash modes
are out of scope.
v3.0.4 attempted a NEON-vectorised compress for aarch64 but the
benchmark regressed 1.5–2× — see the comment on fn compress.
The NEON path is kept under #[cfg(test)] as a cross-check oracle.