Training TF Hub model past 10K steps causes several AUCs to drop

It's highly repeatable, and I suspect that because 10K * batch size is approximately the input size, we are having some bad behavior at the repeat() time. Not sure how to test, maybe with a no-op trainer.

