fix evaluation effiency mismatch between "graph_scoring_efficiency" and "gnn_efficiency_rz"

1. Fix the mismatch between the efficiency value between graph_scoring_efficiency and gnn_efficiency_rz, caused by the former actually plotting the cumulative efficiency and the latter the actual model efficiency.
2. Change the gpus argument in trainer in infer_stage to devices for compatibility with latter versions of pytorch lightning.
3. Delete the shuffle=True setting in dataloader, because it messes up the number of train events when doing inference on multiple GPUs.
4. Add score distribution plot to the ROC curve plot (since these two plots have similar requirements)
5. Some minor changes in plot labels.