Speed up hardlink detection by using a self-sizing hash
table rather than the old linear list search.
On my "hardlink detection torture test", this reduced
user time from 4700 seconds down to 4.2 seconds
and wallclock time from 1:24:48 down to 1:08.
(Yes, that's over one THOUSAND times reduction in user time. ;-)
In the worst case, the new code doubles peak memory usage,
though it could actually reduce memory usage in many cases.