The input density is broadcast at the beginning of an iteration. For this reason we need to broadcast the changed_elements switch in checkMMPmat to consistently reset the mixing_history if needed

Fix for LDA+U linear mixing. I made a mistake when fixing #439. Switched the signs in the wrong place

I encountered a Problem where setting small negative diagonal elements back to 0 lead to much worse convergence (100 instead of 20).
Now the mixing_history is erased if we encounter a value out of the allowed bounds (Only if the density matrix is in the mixer and not mixed linearly)

...as requested by Christoph

Remove -1 for y spin offdiagonal component in greensfunction code since it was seemingly fixed by the noco changes :)

(...probably still not there, but much better)

Add version checks for new optional input parameters (Otherwise even checking if an element is there will result in a invalid xpath error)