ImageNet: spikes in data_time for 16-machine version

[yaroslavvb]

This statistic here spikes to 20ms on 16-machine run:
@bearpelican

cluster/pytorch/training/train_imagenet_nv.py

Line 380 in 13fb813

data_time.update(time.time() - end)

http://35.174.18.63:6006/#scalars&tagFilter=times%2Fdata&runSelectionState=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%3D