Currently, PDF export uses the XMLReader instance created internally by dom4j's SAXReader. As that default XMLReader is configured to remove the entities, the export process has to reset that configuration (and there is little hope that dom4j is going to provide a solution anytime soon).
Instead, I suggest using a XWiki's own XMLReader as delivered by ComponentManager. This has the advantage of not depending on dom4j's decisions, and using the standardized facility that XWiki has for this. I'm preparing a pull request.
Depends on XCOMMONS-2026.