Specialization After Generalization: Towards Understanding Test-Time Training