Superposition disentanglement of neural representations reveals hidden alignment