Hessian Rot/Trans Projection after Mass-weighting

[chselz]

Before the projection of the rot/trans modes were done before the mass weighting. In the dxtb project as well as the gaussian white paper reads as if the correct way to do this is, however, after the massweighting. this lead to problems as no projection was done for linear molecules resulting in wrong frequencies. This patch should elevate this problem, works also for fixed atoms. Testing on the in #693 mentioned linear water molecules yields now two -1802.38 cm^-1 modes. the ordering in the vibspectrum file also seems to be correct now.

This should therefore close #693 .

[lmseidler]

could you explain a bit more: was the problem that no projection was done for linear molecules or that projection before mass weighting changed frequencies significantly in a non-linear case or both? and can you give an example to show both? it might make sense to pin relevant behavior in unit tests. also after merging #1446 this might need some slight updates (there is already a merge conflict)