Wrong Core.Intrinsics.have_fma(Float16) test

[giordano]

This is a known issue, but there's no open ticket about it, so here we go: Core.Intrinsics.have_fma(Float16) always returns false also on hardware which natively supports this datatype:

julia/src/runtime_intrinsics.c

Lines 1759 to 1768 in a5abb6f

	JL_DLLEXPORT jl_value_t *jl_have_fma(jl_value_t *typ)
	{
	JL_TYPECHK(have_fma, datatype, typ); // TODO what about float16/bfloat16?
	if (typ == (jl_value_t*)jl_float32_type)
	return jl_cpu_has_fma(32);
	else if (typ == (jl_value_t*)jl_float64_type)
	return jl_cpu_has_fma(64);
	else
	return jl_false;
	}

The funny result is that, unlike fma, muladd actually uses half-precision fmadd instructions on hardware with native fp16: #57041 (comment)

This issue should hopefully be addressed by #57049.