100 image subset for Decision Boundary attack should be deterministic

[nottombrown]

Currently we shuffle the test dataset to avoid having all birds or all bicycles in a batch.

We also only run the Decision Boundary attack on 100 images because it is too slow to run on the full 1000 image test set.

However, this results in choosing a different 100 images subset each time we run the Decision Boundary attack.

One solution here would be to get the dataset with bird_or_bicyle.get_iterator('test', shuffle=False) and then deterministically shuffle the dataset.