Fix :nth-*(an+b) pseudo-classes selectors

[redapple]

This is a continuation of #42 but following https://www.w3.org/TR/css3-selectors/#structural-pseudos definitions more closely, that is, counting preceding-sibling:: or following-sibling:: explicitly, instead of using context-dependent position():

:nth-child() pseudo-class
The :nth-child(an+b) pseudo-class notation represents an element that has an+b-1 siblings before it in the document tree

In fact, this was suggested in sparklemotion/nokogiri#707 (comment).

It fixes #12, #15, #46.

Another change is xpath_descendant_combinator which now translates to /descendant (instead of /descendant-or-self), which looks more inline with Descendant combinator:

A selector of the form "A B" represents an element B that is an arbitrary descendant of some ancestor element A.

The general case of h2 ~ *:nth-of-type(2) is not fixed by this PR. It looks hard (or impossible) to achieve with XPath 1.0 alone, but it may not be that common case.

[codecov-io]

Current coverage is 95.00%

Merging #60 into master will decrease coverage by 0.04%

@@             master        #60   diff @@
==========================================
  Files             3          3          
  Lines           708        701     -7   
  Methods           0          0          
  Messages          0          0          
  Branches        114        113     -1   
==========================================
- Hits            673        666     -7   
- Misses           19         20     +1   
+ Partials         16         15     -1   

Powered by Codecov. Last updated by b6ba4ff...7fdcf08

[kmike]

Hi @redapple! It seems we'll have hard time finding anyone to review this pull request properly. I'm merging it anyways because

• existing tests pass;
• the PR fixes several issues;
• the approach looks in line with CSS selector specs;
• code looks good.