change the simplex_queue to a SortedKeyList

I suddenly realized the reason why the test_learner_performance_is_invariant_under_scaling was failing for the LearnerND sometimes.

We were saving the rounded loss and then when a subsimplex was created from an existing simplex, we divided the rounded loss instead of dividing the unrounded loss and then rounding it again.

Using a SortedKeyList will make sure this doesn't happen.