Allow choosing between true parallelism and simulated parallelism

Add a config flag that determines whether or not to use true parallelism for running experiments. If run on one machine, federated learning or parallel split learning need postprocessing of the runtime to be comparable to e.g. split learning. Therefore, true parallelism should be distinguished from simulated parallelism. If parallelism is only simulated, the elapsed time (i.e. the wall time) should be logged as well as the maximal time of the individual operations that are supposed to be run in parallel. Then, the parallel time can be determined by postprocessing.