[CostModel] avoid crashing while finding scalarization overhead

The constrained intrinsics have metadata arguments, so the
tests here were crashing as noted in D90554 (and that was
reverted even though this bug exists independently of that
change).