Fix flaky windows test

[nickva]

This PR bypasses the socket tests on windows and fixes a flaky chttpd changes tests where we made a too strong of an assumption about the order in which the shards would be returned.

Thanks to @big-r81 for investigating both issues!

See #4401 (comment) for his windows test failure report.

[big-r81]

Runs through...
See log.

[big-r81]

Running

.\make.cmd eunit apps=chttpd suites=chttpd_socket_buffer_size_test,chttpd_changes_test

Collections
===========

     Total    Fixture       Test      Count     Failed     Errors    Skipped
     11.9s      10.5s       1.4s         40          0          0          0 EUnit        
      0.0s       0.0s       0.0s          0          0          0          0 EXUnit        
      0.0s       0.0s       0.0s          0          0          0          0 JavaScript        
      0.0s       0.0s       0.0s          0          0          0          0 Mango        

Suites
======

     Total    Fixture       Test      Count     Failed     Errors    Skipped
     11.9s      10.5s       1.4s         38          0          0          0 EUnit - chttpd_changes_test
      0.0s       0.0s       0.0s          2          0          0          0 EUnit - chttpd_socket_buffer_size_test

[big-r81]

Would add one additional fixture setup failure...

fixture setup [39,1,1,6]
------------------------

::in function chttpd_security_tests:create_db/1 (test/eunit/chttpd_security_tests.erl, line 60)
in call from chttpd_security_tests:setup/0 (test/eunit/chttpd_security_tests.erl, line 47)
**error:{badmatch,{error,connection_closed}}

[nickva]

::in function chttpd_security_tests:create_db/1 (test/eunit/chttpd_security_tests.erl, line 60)
in call from chttpd_security_tests:setup/0 (test/eunit/chttpd_security_tests.erl, line 47)
**error:{badmatch,{error,connection_closed}}

Hmm, that one is unexpected. We're just trying to create a db with a PUT and we should expect that to always work:

couchdb/src/chttpd/test/eunit/chttpd_security_tests.erl

Line 60 in 23efd8e

{ok, Status, _, _} = test_request:put(Url, [?CONTENT_JSON, ?AUTH], "{}"),

With the small buffer socket test, I can see how connection might be closed suddenly or in a different way on Windows but the simple doc PUT in a security test setup points to something else going on.

Could we be running out of sockets? Not sure descriptor limit a thing on Windows?. Try running just that one security test separately, wonder if that would fail.

Another idea is that there is a genuine race or bug somewhere in couch/ibrowse/mochiweb/erlang vm code that is exposed on Windows only.

[big-r81]

Moved into #4398 ...