Better inlining test in CoreUnfold

[simonpj]

Consider this

f x = let 
        $j y = case x of { True -> e1; False -> e2 }
      in
      case x of 
        True -> ...$j y1...
        False -> ...$j y2...

If a function that scrutinises its argument is applied to a constructor, it becomes much keener to inline. But in this example the function scrutinises a free variable that is evaluated to a known constructor at the call site. At the moment this is ignored, and $j may well not be inlined in situations where it would be jolly good to do so.

This shows up in test perf/should_run/MethSharing where the join points created early only inline because exprIsDupable is just generous enough to do so. If you reduce CoreUtils.dupAppSize by 1, the join point doesn't inline any more. But it should.

The solution is fairly easy: compute discounts on free varaibles as well as arguments. I'm making a ticket because I don't want to execute on this today.

[simonpj]

See #3781, #3755, which would find this optimisation useful.