The DOMFragment token serialization currently stores embedded HTML in data-parsoid.html, which in turn means we need to load data-parsoid (which is present on almost every node) whenever we do a traversal including attribute-embedded HTML.

We *could* move this to a special internal attribute (like data-parsoid-domfragment) but given that data-mw already contains a number of embedded HTML values (like caption) we might as well just add it there. We could split data-mw up further in the future if we wanted found that we were loading too many attributes just looking for embedded HTML, but data-mw is already on fewer nodes that data-parsoid.