Optimize word count

[yufeih]

LoadHtml has huge performance impact, this PR optimizes word count by:

• not load a mutated html
• not split strings

Time to build azure-docs-pr (without git contribution info) drops roughly from 6 mins to 3 mins.

Note:

This implementation is slightly different from the original one, but is more correct:

• Do not special case <title> tag because title, it is already removed in markup phase
• Count both starting tags and ending tags as word separators, the old version only treat ending tags as separator (using string replace hack)

This impacts legacy tests, we could consider removing word_count in as impact test post process.

#2972

[superyyrrzz]

[herohua]