The underlying structures of self-attention: symmetry, directionality, and emergent dynamics in Transformer training