Note [Expanding superclasses]

When we expand superclasses, we use the following algorithm:

transSuperClasses( C tys ) returns the transitive superclasses
                           of (C tys), not including C itself

For example
  class C a b => D a b
  class D b a => C a b

Then
  transSuperClasses( Ord ty )  = [Eq ty]
  transSuperClasses( C ta tb ) = [D tb ta, C tb ta]

Notice that in the recursive-superclass case we include C again at
the end of the chain.  One could exclude C in this case, but
the code is more awkward and there seems no good reason to do so.
(However C.f. GHC.Tc.Solver.Dict.mk_strict_superclasses, which /does/
appear to do so.)

The algorithm is expand( so_far, pred ):

 1. If pred is not a class constraint, return empty set
       Otherwise pred = C ts
 2. If C is in so_far, return empty set (breaks loops)
 3. Find the immediate superclasses constraints of (C ts)
 4. For each such sc_pred, return (sc_pred : expand( so_far+C, D ss )

Notice that

 * With normal Haskell-98 classes, the loop-detector will never bite,
   so we'll get all the superclasses.

 * We need the loop-breaker in case we have UndecidableSuperClasses on

 * Since there is only a finite number of distinct classes, expansion
   must terminate.

 * The loop breaking is a bit conservative. Notably, a tuple class
   could contain many times without threatening termination:
      (Eq a, (Ord a, Ix a))
   And this is try of any class that we can statically guarantee
   as non-recursive (in some sense).  For now, we just make a special
   case for tuples.  Something better would be cool.

See also GHC.Tc.TyCl.Utils.checkClassCycles.

************************************************************************
*                                                                      *
      Classifying types
*                                                                      *
************************************************************************