There may be other io_context instances with child instances, and child
instances with no associated io_context. If we pass 0 to ::waitpid(), we
will reap their processes as well, without updating the state of the
corresponding child instance.
Instead, we call `::waitpid` once for each child we are watching for.
This has some amount of overhead (multiple system calls), but ensures
correct behaviour (providing nothing other than an asio::signal_set is
watching for SIGCHLD).
