Error out on out-of-memory, rather than returning -1, which the sole
existing caller wasn't checking for anyway.  There doesn't seem to be
any use-case for making the caller check for failure here.

Detect failure return from readdir().

Use a less platform-dependent method of calculating the entrysize.
It's possible, but not yet confirmed, that this explains bug #6733,
in which Mike Wilson reports a pg_upgrade crash that did not occur
in 9.1.  (Note that load_directory is effectively new code in 9.2,
at least on platforms that have scandir().)

Fix up comments, avoid uselessly using two counters, reduce the number
of realloc calls to something sane.