On the Hardness of Probabilistic Neurosymbolic Learning