keccak: avoid materializing round constants array

[shuklaayush]

• Switch from eager array .map() to lazy iterator mapping for round constants
• Avoid materializing a temporary [L; ROUNDS], reducing stack usage

[newpavlov]

With enabled optimization this code gets optimized into a pre-computed const table, so this change is not necessary.

[shuklaayush]

@newpavlov thanks for the reply! the conversions are constant-folded into a table, but the owned array is still materialized in the optimized aarch64 codegen

before:

sub sp, sp, #0x190

ldp q0, q1, [x8, #0x80]
stp q0, q1, [sp, #0xf0]
ldp q0, q1, [x8, #0xa0]
stp q0, q1, [sp, #0x110]
ldp q0, q1, [x8, #0x40]
stp q0, q1, [sp, #0xb0]
ldp q0, q1, [x8, #0x60]
stp q0, q1, [sp, #0xd0]
ldp q0, q1, [x8]
stp q0, q1, [sp, #0x70]
ldp q0, q1, [x8, #0x20]
stp q0, q1, [sp, #0x90]

after:

sub sp, sp, #0xc0

adrp x9, ...
add  x9, x9, ...
ldr  x9, [x9, x8]

on my macbook air m3, running

RUSTFLAGS='--cfg keccak_backend="soft"' \
cargo +nightly bench -p keccak --bench mod -- keccak_f1600 --exact

goes from 172.63 ns before to 145.93 ns after

[newpavlov]

Interesting. I opened #129 to resolve it in a different and hopefully less fragile fashion.