Noticeable performance impact of change to handle peer disconnects

Hello!

I tried porting an existing httpaf test to httpun for the TechEmpower web framework benchmark and [posted about it here](https://discuss.ocaml.org/t/lwt-multi-processing-much-more-performant-than-eio-multi-core/16395/11). It looks like the change that was introduced [here](https://github.com/anmonteiro/gluten/pull/60/files#diff-d2f6985412478e50917d9d83c6a35b2322073e94f65066a7232e1418ab14e7cdR90-R95) actually carries a ~15% performance overhead. Not sure if there's a better way to deal with this but thought it was worth mentioning.