Toffoli and Hadamard are computationally universal -- that is, they can be used to carry out any quantum computation. However, they do so by implementing quantum gates in an encoded way. Indeed, this is necessary since both Toffoli and Hadamard have only real entries, so there is no way to obtain quantum gates with complex entries, unless one uses some encoding (see the paper you linked). That means that Toffoli and Hadamard are not universal in the sense that you can use them to construct any gate. In particular, there is no way to actually construct the $\pi/8$ or the Deutsch gate (except for special angles), or to even approximate them.