Simplifying Distributed Neural Network Training on Massive Graphs: Randomized Partitions Improve Model Aggregation