On some platforms, vec4 support has obvious perf advantage over plain fp32. We need to enable this in TVM.