Slow compile times

Compiling `tokenizers` takes around 35 seconds in release mode.
Most of this (25s) is spend compiling `onig_sys`.

An alternative to using `onig_sys` would be appreciated. This crate already uses the `regex-*` ecosystem, so if that could be used, it would be a big win for compile times and binary sizes.

Otherwise, if you point me to where other regex libraries fall down, I'm happy to look into alternatives.