You must call destroylock before freeing the memory holding lockstat
enabled lock.  I did this for most of the locks.

I might have broken something, but gc threads were running too
infrequently for testing.  A gc thread might run at a higher rate due
to the heuristic in gc_end_epoch.

- working crange_search
- crange_search takes sz as argument
- better for deciding when a node can be delayed freed