contification transformation

[Matthew Fluet]

> 1. A/ancestor analysis + old transformation
> 2. A*/parent analysis + new transformation
...
> Assuming it's not noticeably slower, I lean slightly towards (2), but mostly
> just because it's already implemented. 

I'll run some real tests of the timing between the two transformations,
but I don't think it will be significant.  Really, the only additional
cost should be deciding the order in which to process the nodes.  After
that, everything is pretty much the same; the only additional overhead is
dealing with problematic mutual recursion, which in practice is almost
non-existent.