Use <b> elements to identify probable good first paragraphs
Improves upon our heuristic for identifying the true lead paragraph for
extracting lead introductions. The original version was unreliable on
non-English wikis, particularly dewiki; and the updated version in commit
bee6c764 was overly aggressive and had significant problems on itwiki.
With this change the results as tested by compare-extracts.js are
consistently very good across wikis.