Optimize Lwt_io.establish_server(_with_client_socket)

[aantron]

Found by @bluddy in inhabitedtype/httpaf#54 (comment):

OK, I redid the test with the single yield () modification above. For the record, it involves turning line 1608 in lwt_io.ml to

 Lwt_main.yield () >>= fun () -> accept_loop ()

This results in a huge latency improvement for the HTTP server benchmark being discussed in inhabitedtype/httpaf#54 (comment); see surrounding comments for details. However, it also decreases throughput.

There may be some variant of this optimization that gives both low latency and high throughput. There is a risk of over-optimizing for the specific benchmark, however.

One suggestion is to take the accept_n branch, and yield between every N accepts.