Signal -k children individually, not the process group

[MRX-72]

Ncat in listen mode forwards a termination signal with kill(0, ...), which
targets its whole process group. Ncat does not create a group of its own, so
that group is the one it inherited from whatever launched it, and the caller
is terminated as well. The handlers are installed for listeners that do not
use -k either.

Reported against the OVN system test suite, where the cleanup SIGTERM to
ncat -4 -l -u -p 1234 -d 0.1 -c cat killed the test runner and the suite
exited with status 143.

Track the PIDs that are forked for connections and signal those one by one.
Slots are released when a child is reaped, and SIGCHLD is blocked while the
signal is forwarded so that a PID cannot be recycled in the middle of it.
waitpid() with WNOHANG confirms each PID is still one of our children before
it is signalled. The children restore the default dispositions for these
signals, so that they terminate instead of forwarding the signal again.

Fixes #3489

[MRX-72]

This has landed in master as 4d67c06 — thanks @bonsaiviking.

Noting it here since the SVN-side commit doesn't auto-close the PR, and the diff is now empty against master apart from your follow-up in 469fff8: exit() → _exit() inside the handler. Good catch, and it's the right call — exit() runs atexit handlers and flushes stdio, neither of which is async-signal-safe, so it could deadlock against an interrupted stdio lock. That exit() predates this change; I moved the call without questioning it.

Happy for this to be closed whenever suits.

[dmiller]

Your 'thanks' attributed the wrong person, in case you'd like to acknowledge the intended person.