Virtual files are dumped to disk at each iteration in MPI processes, drastically reducing memory usage

move binary file output to virtual file passed via MPI to process 0, thread 0 for actual writing to disk

put the preprocessor checks for logger entries of different matrix-inversion options in the correct order

- fix cluster (make logging message buffers omp local)
- add cluster test case 3

- fix logger race in omp threads
- add mpi to docker images