Improve performance of differential equation

[JordiBolibar]

There might still be room for improvement in terms of speed for the rotation differential equation:

    function ude_rotation!(du, u, p, t)
        # Angular momentum given by network prediction
        L = U([t], p, st)[1]
        du .= cross(L, u)
    end

L is currently out-of-place. I'll try to implement it in-place and see if the reverse differentiation packages don't complain about that.