Since doc.getElementsByTagName("font") returns live array, _forEachNode
must be called with backward=true or it will skip elements.

Also, _forEachNode with backward=true was NYI.

1) Avoid conversion of whitespace text nodes into paragraphs. They
create a lot of noise and actually prevent sibling joining logic from
working in many pages.

2) Handle case when adjacent content is actually located in parent's
sibling node instead of top candidate’s sibling.

Solution: strip one level of empty <DIV> elements so they don’t
obstruct merging adjacent content downstream.