Note [Case to let transformation]

If a case over a lifted type has a single alternative, and is being
used as a strict 'let' (all isDeadBinder bndrs), we may want to do
this transformation:

    case e of r       ===>   let r = e in ...r...
      _ -> ...r...

We treat the unlifted and lifted cases separately:

* Unlifted case: 'e' satisfies exprOkForSpeculation
  (ok-for-spec is needed to satisfy the let-can-float invariant).
  This turns     case a +# b of r -> ...r...
  into           let r = a +# b in ...r...
  and thence     .....(a +# b)....

  However, if we have
      case indexArray# a i of r -> ...r...
  we might like to do the same, and inline the (indexArray# a i).
  But indexArray# is not okForSpeculation, so we don't build a let
  in rebuildCase (lest it get floated *out*), so the inlining doesn't
  happen either.  Annoying.

* Lifted case: we need to be sure that the expression is already
  evaluated (exprIsHNF).  If it's not already evaluated
      - we risk losing exceptions, divergence or
        user-specified thunk-forcing
      - even if 'e' is guaranteed to converge, we don't want to
        create a thunk (call by need) instead of evaluating it
        right away (call by value)

  However, we can turn the case into a /strict/ let if the 'r' is
  used strictly in the body.  Then we won't lose divergence; and
  we won't build a thunk because the let is strict.
  See also Note [Case-to-let for strictly-used binders]