Keep a set of the already read page tree nodes and each time we follow a new one check if we already processed it so we don't end in a infinite loop.
Fixes crash in malicious pdf found at <a href="http://projects.info-pull.com/moab/MOAB-06-01-2007.html">http://projects.info-pull.com/moab/MOAB-06-01-2007.html</a>;

It may seem quite intensive but my measurements indicate the page tree processing is I/O bound as i got the same average times with and without this patch. Obviously if anyone wants to measure more it'll be appreciated.