Revisit tokenizer microoptimizations

[kmcallister]

For example, there's an old branch to use SSE 4.2 string instructions in tokenization, which realized a speedup of about 10% and could probably be pushed much further.

[kmcallister]

https://crates.io/crates/llvmint has the necessary intrinsics. That might be a nicer route than inline assembly, even though it's still not portable.