PackInfo code generation

[jngrad]

Fixes #4921

Description of changes:

• reduce volume of the LB communication during the streaming step
• avoid superfluous LB ghost communication outside the streaming step
• use AVX streaming kernels

[RudolfWeeber]

On my system, the LB benchmark with 1k particles/core is now 30% faster for 1 and 8 cores and for GPU. That's a significant imporement!

For grids below ~40 cubed, the CPU is version is now close to the GPu one (30% slower, or so). At some point, we might look into that some more. Long term, w emight

• split the particle coupling to hide the laency of getting the interpolated velocities form the GPU
• and/or move the entire particle coupling to the GPU