Fix bug by adding cudaDeviceSynchronize()

The scaling coarse to fine for bulk nodes should not overlap with the collision kernel for boder nodes on the fine level .

(cherry picked from commit dafce4df)