fix discrete divergence

We want to use the divergence with zero boundary only in the case of dd.
It actually didn't matter for chambolle, since that boundary information
is never propagated through the algorithm for initialization with 0, but
this way is more appropriate.