Implement axy using sparseRangeFor()

This way it works for multi-type matrices again.
This helps to fix the dune-tnnmg test which currently
breaks.

Notice that this may be slightly less efficient, because
we use a new temporary for each loop instance. This is
necessary because we may need a different type each time.
This may be fixed by adding a specialization for the
case that the corresponding vector is a classic dynamic
container.