Walberla performance tracking ticket

[RudolfWeeber]

• more selective ghost communication LB: avoid double communicaiton #4921
• Use AVX vectorization in all kernels (streaming, boundaries, reset force, ...). (It might be useful to automate the generatoin of kernel_traits.hpp, first (Simplify work with walberla kernels #4988)
• replace
• switch to pull scheme and use combined stream-collide sweep (postponed, as I couldn't get our patched bounce back boundary condiont working with pull scheme)
• Generate the packinfos using PyStencils and use vectorization (currently, we use hand-written pack infos form Walberla propper). In particular, do such generated packinfos outperform the currenty used memcpy pack info on the GPU
• Multi-GPu simulatoin with cCuda-aware MPI

[RudolfWeeber]

Just realized that we did not yet vectorize the streaming kernel. This is probably a relatively low hanging fruit for performance on CPU.