This is a pretty big patch, but which substantially iproves the subsumption
check.  Trac #9569 was the presenting example, showing how type inference could
depend rather delicately on eta expansion.  But there are other less exotic
examples; see Note [Co/contra-variance of subsumption checking] in TcUnify.

The driving change is to TcUnify.tcSubType.  But also

* HsWrapper gets a new constructor WpFun, which behaves very like CoFun:
       if     wrap1 :: exp_arg <= act_arg
              wrap2 :: act_res <= exp_res
       then   WpFun wrap1 wrap2 : (act_arg -> arg_res) <= (exp_arg -> exp_res)

* I generalised TcExp.tcApp to call tcSubType on the result,
  rather than tcUnifyType.  I think this just makes it consistent
  with everything else, notably tcWrapResult.

As usual I ended up doing some follow-on refactoring

* AmbigOrigin is gone (in favour of TypeEqOrigin)
* Combined BindPatSigCtxt and PatSigCxt into one
* Improved a bit of error message generation