Note [Local instances and incoherence]

Consider
   f :: forall b c. (Eq b, forall a. Eq a => Eq (c a))
                 => c b -> Bool
   f x = x==x

We get [W] Eq (c b), and we must use the local instance to solve it.

BUT that wanted also unifies with the top-level Eq [a] instance,
and Eq (Maybe a) etc.  We want the local instance to "win", otherwise
we can't solve the wanted at all.  So we mark it as Incohherent.
According to Note [Rules for instance lookup] in GHC.Core.InstEnv, that'll
make it win even if there are other instances that unify.

Moreover this is not a hack!  The evidence for this local instance
will be constructed by GHC at a call site... from the very instances
that unify with it here.  It is not like an incoherent user-written
instance which might have utterly different behaviour.

Consider  f :: Eq a => blah.  If we have [W] Eq a, we certainly
get it from the Eq a context, without worrying that there are
lots of top-level instances that unify with [W] Eq a!  We'll use
those instances to build evidence to pass to f. That's just the
nullary case of what's happening here.