LB coupling for RegularDecomposition without ghost shifts.

[pkreissl]

Fixes #4466, fixes #4432, closes #4440

Description of changes:

• ports n-square lb coupling support from the walberla branch back into the python branch
• merges PR WIP: Do not shift ghost positions in domain decomposition #4440 and fixes CI issues along the way
• expose some funtions in lb_particle_coupling.hpp to get rid of code duplication in (the fixed version of) lb_inertialess_tracers.cpp

[RudolfWeeber]

@pkreissl, what's the status. Given that this Pr is high-risk and the feature freeze is fast aproaching.

[pkreissl]

Resolving the code duplication in the LB coupling of both, regular particles and inertialess tracers, proved to be not quite as straight-forward as anticipated. I am open for suggestions for improvement.

• In lb_inertialess_tracers: former in_local_domain() was basically the same as in_local_halo() of lb_particle_coupling.cpp
• In the current commit inline in_local_domain() is defined in both .cpp files. Putting it into the lb_particle_coupling.hpp is not directly possible, as the function requires an include "grid.hpp". This would lead to an indirect include of mpi.h in electrokinetics_cuda.cu, which is not permissible.

@RudolfWeeber , can you please also have an extra-close look at the coupling logic in the inertialess tracer coupling -- I think I got it right, and the tests pass, but still...

[jngrad]

From a C++ perspective, this looks good to me. I ran the lb_momentum_conservation.py test on CPU with particle coupling, kT=0, 500 time steps, random initial positions and velocities for 20 particles on MPI topologies [1 1 1], [2 1 1], [3 1 1], [2 2 1], [1 2 2], [4 1 1]. The particle forces at the end of the simulations were identical within 7 decimal places, and were also identical to the same simulation on 1 MPI node with the N-square cell system.

[pkreissl]

I think @christophlohrmann also wanted to have a look at this PR. What is the status?

[pkreissl]

I just also ran the following tests on 1, 2, 3, 4, and 6 cores, respectively, to run them with different MPI topologies:

• engine_lb.py as additional test for active particles.
• virtual_sites_tracers.py; for that, I had to adapt virtual_sites_tracers_common.py, i.e., I used box_l = [6, 6, 12] and increased the number of integration steps for establishing the steady state flow field in the advection test (line 81) to 800.

All these additional test passed.