intel/compiler: assert restrictions on conversions to half-float