ballet: speedup AVX512 reedsol encode

[drubin-fd]

Zen5:
Before (GCC):

average time per encode call 1220 ns

After (GCC):

average time per encode call 870 ns

This is a 30% speedup, when measured on a Zen 5 machine. The main benefit of the newer approach is that it actually uses AVX512, where the old one was only using AVX2. The new version is also written in Zig, with the generated assembly checked in. This makes auditing the implementation and understanding it much easier.

Icelake:
Before (Clang):

average time per encode call 2049ns

After (Clang):

average time per encode call 1745ns

Icelake benchmarks are more interesting, and more accurate, as LLVM does not yet have scheduling information for Zen5, which makes it schedule somewhat poorly (we could win with more carefully crafted asm that uses zmm registers).