[GNN] Acc eval script change: incompatibility or corrupted file?

After the change to avoid using memmap when loading the labels #2081, we're encountering pickling errors via numpy.
We're unable to load the labels with numpy.load. The previous loading via np.memmap (https://github.com/mlcommons/inference/commit/be6ff52235b74a2f4ef85bf86fa3785045229fa8) did work on our environments without issues

Steps to reproduce:

1. wget -c https://igb-public.s3.us-east-2.amazonaws.com/IGBH/processed/paper/node_label_2K.npy
2. python -c "import numpy as np; labels = np.load('node_label_2K.npy', mmap_mode=None)"
3. Error: ValueError: Cannot load file containing pickled data when allow_pickle=False

After setting allow_pickle=True, we face another error: _pickle.UnpicklingError: invalid load key, '\x00'

Packages used: numpy==1.26.4, torch==2.1.0+cpu

Curious if there's a dependency issue on my side.
cc: @arjunsuresh @nv-alicheng

Was the previous mode of loading the labels incorrect leading to incorrect labels? Or the current modification fixes a previously missed bug? If the latest change doesn't concern correctness of loaded labels, can we bring back the old way so as the default, and switch to np.load if --no-memmap is passed?

Start by reproducing the failure with the wget and Python commands in the issue, then inspect the label-loading change from #2081 and commit be6ff52235b74a2f4ef85bf86fa3785045229fa8. Compare the downloaded node_label_2K.npy format with the loading modes used by the GNN accuracy evaluation. Done means identifying whether the file or loader is incompatible and establishing a verified, documented loading behavior.