Adjust handleCopyOut() to stop trying to write data once it's failed
one time.  For typical cases such as out-of-disk-space or broken-pipe,
additional attempts aren't going to do anything but waste time, and
in any case clean truncation of the output seems like a better behavior
than randomly dropping blocks in the middle.

Also remove dubious (and misleadingly documented) attempt to force our way
out of COPY_OUT state if libpq didn't do that.  If we did have a situation
like that, it'd be a bug in libpq and would be better fixed there, IMO.
We can hope that commit fa4440f5 took care
of any such problems, anyway.

Also fix longstanding bug in handleCopyIn(): PQputCopyEnd() only supports
a non-null errormsg parameter in protocol version 3, and will actively
fail if one is passed in version 2.  This would've made our attempts
to get out of COPY_IN state after a failure into infinite loops when
talking to pre-7.4 servers.

Back-patch the COPY_OUT state change business back to 9.2 where it was
introduced, and the other two fixes into all supported branches.