Memsync

[Thinkpiet]

• Adds explicit memory synchronization (device to/from host) to CUDA version of SimpleMD, based on state machine (see MoleculeContainer::synchronizeMemory)
• Adds various performance optimizations and improvements
• SimpleMD GPU performance drastically improved!

(TODO) boundaryTreatment needs refactoring before this can be merged to master. See simplemd/MolecularDynamicsSimulation.cpp:534

• putBoundaryParticlesToInnerCellsAndFillBoundaryCells: Separate into independent steps, do not call them if not required (no parallel boundary / no periodic boundary)
• emptyGhostBoundaryCells: Merge separate kernels calls / 6 domain boundaries into one; do not call if not required (no parallel boundary / no periodic boundary)
• Make sure that re-enabling _boundaryTreatment in MolecularDynamicsSimulation::simulateOneTimestep does not affect run time of ./benchmarks MD120 (it should not consume runtime as it does not perform any operations in a reflecting 1-rank scenario)