The original coding in commit 1e8a8500 didn't use PQconnectPoll per
spec, and while the rewrite in e434ad39 is closer, it still doesn't
guarantee to wait until the socket is read-ready or write-ready (as
appropriate) before calling PQconnectPoll.  It's not clear whether
that omission is causing the continuing failures on buildfarm member
bowerbird; but given the lack of other explanations meeting the
available facts, let's tighten that up and see what happens.

An independent issue in the same loop was that it had a race condition
whereby it could clear the process's latch without having serviced an
interrupt request, causing failure to respond to a cancel while waiting
for connection (the very problem 1e8a8500 was meant to fix).

Discussion: https://postgr.es/m/7295.1489596949@sss.pgh.pa.us