Parallelizing chains with custom likelihood on multiple cores

Initially I thought that this is the best explanation, but I agree that for the example I posted above this does not work. If it would be properly parallel, the sampling time would not change as cores and chains increase equally.