We should create a benchmark that focuses on number parsing... https://github.com/simdjson/simdjson/discussions/2185