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 520f1db1)

SetRecvFsPostDev27() is now called after communication in all directions is finished. This is needed for communication hiding, when cutting the domain in multiple directions.

Avoid calling method to find neighborNodes in every timestep

cherry-picked from d954628

also rename methods exchangePostCollDataGPU27() to exchangeCollDataGPU27()

* move cudaEventRecord to UpdateGrid27
* use variables for stream indices in UpdateGrid27
* create and destroy streams in Simulation.init() and Simulation.free()

Previously, the event was also triggered when data was written. This behavior caused crashes of VF