Completely revamp node binding for AST matchers.

This is in preparation for the backwards references to bound
nodes, which will expose a lot more about how matches occur.  Main
changes:
- instead of building the tree of bound nodes, we build a "set" of bound
  nodes and explode all possible match combinations while running
  through the matchers; this will allow us to also implement matchers
  that filter down the current set of matches, like "equalsBoundNode"
- take the set of bound nodes at the start of the match into
  consideration when doing memoization; as part of that, reevaluated
  that memoization gives us benefits that are large enough (it still
  does - the effect on common match patterns is up to an order of
  magnitude)
- reset the bound nodes when a node does not match, thus never leaking
  information from partial sub-matcher matches for failing matchers

Effects:
- we can now correctly "explode" combinatorial matches, for example:
  allOf(forEachDescendant(...bind("a")),
  forEachDescendant(...bind("b"))) will now trigger matches for all
  combinations of matching "a" and "b"s.
- we now never expose bound nodes from partial matches in matchers that
  did not match in the end - this fixes a long-standing issue

FIXMEs:
- rename BoundNodesTreeBuilder to BoundNodesBuilder or
  BoundNodesSetBuilder, as we don't build a tree any more; this is out
  of scope for this change, though
- we're seeing some performance regressions (around 10%), but I expect
  some performance tuning will get that back, and it's easily worth
  the increase in expressiveness for now

llvm-svn: 184313