Fix deterministc group_shuffle_split

[nilsleh]

Sets are unordered and therefore, repeated calls were yielding different train and val sizes for Cyclone dataset.

[adamjstewart]

Let's find a way to test this so the same bug doesn't happen again.

[isaaccorley]

@nilsleh Would love to get this in the 0.5.2 release!

[nilsleh]

@nilsleh Would love to get this in the 0.5.2 release!

Thanks for the reminder :)

[nilsleh]

@adamjstewart and @isaaccorley not sure how to simulate repeated calls to the function after restarting script/kernel, so I thought separate processes might be a way to go, but actually not sure

[adamjstewart]

@adamjstewart and @isaaccorley not sure how to simulate repeated calls to the function after restarting script/kernel, so I thought separate processes might be a way to go, but actually not sure

This feels like overkill. Depending on the size of our fake dataset, can we just run the test once, print the order, then hardcode that in the test code? As long as it is always the same, it's deterministic.