It allows Starman to sit nicely behind a balancer keeping backend connections
open for a long time (pre-opening or just using keep-alive). Without
this change the behaviour is 5 seconds timeout, and depending on backend
utilization it could be likely that a new request hit the timeout just after
a client connection comes in.
t=0.0s, balancer opens connection, Starman block reading headers.
t~4.9s, client connection hits balancer, balancer assign preopened
t~5.0s, RST if race. or FIN to balancer with request not processed.