-------------------------
	Remove function coercions
	-------------------------

    (coerce (T1->T2) (S1->S2) F) E
  ===>
    coerce T2 S2 (F (coerce S1 T1 E))

This is a generally good transformation, but it
still doesn't solve the problem I was after.  Consider


  newtype T = MkT (Int -> Int)

  p :: T->T;   p = ...
  q :: T;      q = ...

  foo :: T
  {-# INLINE foo #-}
  foo = p $ q

  f = \y -> ...((coerce (Int->Int) foo) 3)...


Trouble is, foo doesn't see the argument because of the coerce, so it
thinks it's a lone variable and doesn't inline.

Another problem is that since $ ins't inlined into foo's RHS, foo
looks like a redex, which we are reluctant to inline inside a lambda,
even with an INLINE pragma.  Maybe we should be bolder?

Anyway, this commit is an improvement to Simplify, but the story is not
over!