Use AVX-512 when available

[Shnatsel]

On Zen 4 this increases the improvement over the base64 crate from 3.5x to 4.5x

[Shnatsel]

I haven't tested this (other than benchmarking) so it is entirely possible that the handling of the remainder needs to be changed too.

[Shnatsel]

Actually, I've re-measured and it seems to be a measurement error on my part. Both the avx2 and avx512 modes produce the exact same performance.

It seems that the avx512 target feature is not yet stable, and is ignored: rust-lang/rust#44839
Or maybe it's just the Zen4 double-pumped implementation of AVX-512 not being any different than the AVX2. I'd have to look at the disassembly to confirm. I'll close this until then.