intel/compiler: handle extended math restrictions for half-float