Transformer Normalisation Layers and the Independence of Semantic Subspaces