thread-safety: strerror_r()

There are only a few (not necessarily thread-safe) strerror() calls in 
the backend; most other potential users use %m in a format string.

In two cases, the reason for using strerror() was that we needed to 
print the error message twice, and so errno has to be reset for the 
second time.  And/or some of this code is from before snprintf() gained 
%m support.  This can easily be simplified now.

The other is a workaround for OpenSSL that we have already handled in an 
equivalent way in libpq.

(And there is one in postmaster.c, but that one is before forking.)

I think we can apply these patches now to check this off the list of 
not-thread-safe functions to check.