Fix low-probability bug in relcache startup: write_irels wrote the

pg_internal.init file in-place, which meant that if another backend
started at about the same time, it might read the incomplete file.
init_irels tries to guard against that, but I have now seen a crash
due to reading bad data from a partly-written file.  (This may indicate
a kernel bug on my platform?  Not sure.)  Anyway, clearly the safest
course is to write the new pg_internal.init file under a unique temporary
filename, and rename it into place only after it's all written.