[tfel-math] Add better support class template argument deduction (CTAD)

[thelfer]

C++-17 introduces class template argument deduction CTAD, which was never supported in tfel.

Basically, the following code shall work:

stensor a = {1, 2, 3}; // a is an stensor<1u, int>
tensor b = {1, 2, 3 , 4 , 5} // b is a tensor<2u, int>