Mind the truncation gap: challenges of learning on dynamic graphs with recurrent architectures