Synergies between Disentanglement and Sparsity: Generalization and Identifiability in Multi-Task Learning