This PR changes one of the haveToRemove checks to allow for short paragraphs, such as the written dialog in the linked issue, by adding linkDensity to the primary "short content" check. The rationale being that short strings which don't contain any links are likely to be text the user would want to read, provided that the initial preprocessing has already removed the bulk of the page elements.

I regenerated all test cases to check for regressions and added a couple of new checks to deal with those:

• adWords and loadingWords regexes to help identify ad blocks and loading indicators.
• textDensity + image count to exclude elements without useful content.

Test case changes:

• citylab-1 now includes the published time in the Readability content.
• ehow-1 lost an extraneous "Other People Are Reading" header but gained an extraneous "Found This Helpful" header
• ehow-2 lost its "Other People Are Reading" header and gained the word "Save" previously used for a button
• engadget now shows the base price and score originally present in the product review
• firefox-nightly-blog no longer shows "Leave a Reply"
• mercurial now correctly shows a previously excluded code snippets and commands
• qq now shows the page header with published date
• toc-missing now incorrectly shows "Interactive Editor"
• wikipedia now shows image captions wrapped in <div><p>...</p></div>

The changes which are definitely regressions seem like an acceptable trade for the improvements gained. I would like some input on whether the adWords and loadingWords regexes are acceptable though. They are scoped so that they will only produce a match against a node's entire innerText string, so it's unlikely to impact real content, but it still feels like a slippery slope.

Closes #861