Spectral Editing of Activations for Large Language Model Alignment