test: improve cluster-disconnect-handles test

[mscdex]

This commit fixes two issues in test-cluster-disconnect-handles:

1. If the master's TCP connection to the worker fails, the worker process stays alive and causes many other tests that use the same common port number to also fail (with EADDRINUSE).
2. One particular problem that can cause the master's TCP connection to fail is attempting an IPv6 connection to the worker when no IPv6 network interfaces are available (this was the case on my local Linux machine).

[mscdex]

CI is all green except unrelated errors on Windows.

[cjihrig]

Hmm... is this the only test that suffers from this problem?

[mscdex]

I don't know, but this was the only one I personally saw that was causing this problem.

[cjihrig]

LGTM with a question and comment.

[cjihrig]

Does this close #3907? There is also #4009 for that test.

[mscdex]

@cjihrig I don't know, that might be a separate issue.

This is what I get when I don't have the IPv6 check in place:

Debugger listening on port 12946
Debugger listening on port 12346
/home/mscdex/git/node/test/parallel/test-cluster-disconnect-handles.js:72
throw ex;
^
Error: connect EADDRNOTAVAIL :::12346 - Local (:::0)
at Object.exports._errnoException (util.js:873:11)
at exports._exceptionWithHostPort (util.js:896:20)
at connect (net.js:844:14)
at net.js:942:9
at doNTCallback0 (node.js:430:9)
at process._tickDomainCallback (node.js:400:13)

[cjihrig]

Does listening on 'localhost' work for you?

[mscdex]

It would probably work for me because I only have localhost resolving to 127.0.0.1, but by default most modern Linux distros have localhost resolve to both IPv4 and IPv6 addresses (even if there are no IPv6 interfaces).

[cjihrig]

OK, I don't want to hold up this PR, I was just looking for a cleaner alternative.

[mscdex]

Landed in 49b5445.