In other words, the PVP is not rigorously specified, and this particular question is uncharted territory.
My initial interpretation of the PVP is one of subtyping. I consider loosening type signatures as a minor version bump, because the new thing still fits into holes shaped like the old thing.
However, seen a different way, you could say that previously, IO () was just a value, but now, it's (forall a. IO a), which is basically (a:Type -> IO a), if Haskell were dependently typed. Going from a value to a function is definitely a major change.
Was that sufficiently unhelpful? :P I think my conclusion is perhaps that the PVP should remain undefined on this particular question, since I see no obvious answer. Or, since the PVP already encourages a cautious over-estimate of what constitutes a breaking change, perhaps we should go with this fallback: "when in doubt, call it a major change."