scramble input

[mortonjt]

If the input tables are sorted / clustered beforehand -- the algorithm will drastically underperform.

There needs to be some scrambling beforehand -- preferably at each iteration. See torch dataloader: https://pytorch.org/docs/stable/data.html