The readme says that encoding is 20% faster when reusing an Encoder instance. If that's the case, is there any reason not to just memoize them in encode (or at least reuse a default instance when no options are set)?