TcInstDcls.tcSuperClasses was getting increasingly baroque as a
succession of tickets (#10423 being the latest) pointed out that
my cunning plan was not so cunning.

The big issue is how to restrict the evidence that we generate
for superclass constraints in an instance declaration to avoid
superclass loops.  See Note [Recursive superclasses] in TcInstDcls
which explains the plan.

The question is how to implement the plan.  The new implementation is
much neater, and is described in Note [Solving superclass constraints]
in TcInstDcls.