How can one ensure that indexing the h node matrix with the topk scores applies to each member of a batch of graph data?