Toward Efficient Robust Training against Union of \ell_p Threat Models