More nesting plumbing. Still does nothing because we don't parse the
nested rules.

Should be trivial to prove this patch doesn't change any behavior so
far, but I want to land it on its own because it can have performance
implications.

This follows the pattern of what we do with other rules like layers and
container conditions, that is, keep the ancestor selectors in a stack,
and poke at the last one in order to replace the ancestor.

This changes the behavior of replace_parent_selector as with the newer
version of the spec, stuff like:

div {
  .foo {
    stuff
  }
}

Should work as div .foo. A test is added for this case.