Tensor product kernels: specialize a function for complex numbers

[kronbichler]

I realized that the code generated by compilers for the matrix-free tensor product kernels when used for complex numbers is rather poor, because it will not exploit fused multiply-add functions in the inner reduction loops and thus performs unnecessary work. This is easy to fix, especially with if constexpr facilities: We write the code in a way that lets the compiler straight-forwardly use FMA operations on both the real and imaginary part.