This doesn't fully work, in that more test cases are broken. However,
this is mostly due to questionable inference rules in the past, and
so fixing these errors should lead to a more correct inference algorithm.
(i.e. we need to handle things like (List X ... a) (List Y .. b),
where a and b are not the same bound. We'd started this work before,
but never actually gone through with it, since smashing regular and
dotted type variables into the same environment meant some things
magically worked when they possibly shouldn't have.)
b5d4d54d45