Wide Graph Neural Networks: Aggregation Provably Leads to Exponentially Trainability Loss