New method of whitespace joining

[jemrobinson]

Rewrite of the plain_html logic. The sequence is now as follows:

• Insert space into non-spaced comments so that html5lib can interpret them correctly
• Convert the HTML into a Soup parse tree
• Remove comments and DOCTYPE strings
• Process CDATA (currently we remove it)
• Strip tag attributes apart from 'class' and 'style'
• Remove blacklisted elements
• Unwrap elements where we want to keep the text but drop the containing tag
• Process elements with special innerText handling
• Process unknown elements
• Consolidate text, joining any consecutive NavigableStrings together
• Remove empty strings and elements
• Replace <br> and <hr> elements with paragraph breaks, wrapping in <p> tags where needed
• Normalise all strings, removing whitespace and fixing unicode issues
• Wrap any remaining bare text in a suitable block level element
• Recursively replace any elements which have no children or only zero-length children

The main differences are

1. Move the string consolidation before whitespace normalisation (which closes #45 and closes #46).
2. Combine the search and replacement of <br> and <hr> elements into one function (removing fragile function ordering problems)

[jemrobinson]

I've added some tests which fail if the ordering constraints are violated. The paragraph breaking code calls consolidate in between the identify and split stages, so that the newly-added break indicators can be included in the correct string.

[jemrobinson]

Reviewed in person with @martintoreilly . Only outstanding question is which of these is correct behaviour:

1. <div>text</div> => <div>text</div>
2. <div>text</div> => <div><p>text</p></div>

I think it is (1) and this is what the code currently does, but should be confirmed.

Confirmed that (1) is correct with @martintoreilly . See comment in plain_html.py:

We do this to ensure that there is a strong, unique correspondance between presentational paragraphs and DOM structure
     - all presentational paragraphs should be the only content associated with their immediate parent
     - all presentational paragraphs at the same conceptual level should be equally nested
     - the string as displayed in the browser should be equivalent to the innerHTML of the parent (so that indexing is equivalent between presentation and source)

    The following examples should not be allowed:

     1. Two presentational elements at the same DOM level have non-equivalent index levels
       <div index="1.1">
         text
         <p index="1.1.1">more text</p>
       </div>

     2. Index 1.1 might contain both strings
       <div index="1.1">
         <p index="1.1.1">more text</p>
         text
       </div>

     3. Two presentational paragraphs are included in the same index
       <div index="1.1">
         text
         <p index="1.1.1">more text</p>
         yet more text
       </div>