Emergent Interpretable Symbols and Content-Style Disentanglement via Variance-Invariance Constraints