skip assembly of overlap cells in FastDG assembler

When assembling a parallel problem, we currently need one layer of overlap to correctly assemble skeleton integrals on the processor boundaries, but we don't really have to do any calculations on those overlap cells.

This merge requests changes the FastDG grid operator to skip those cells. One small consequence of this change is that some intersections will now be visited from the opposite side as before (if the adjacent cell with the higher index is an overlap cell).