Improve optimistic selection

[ch3njust1n]

threshold for forming a hyperdge. if all peers in party are already at good points in space, then no need to form an edge and train, which would result in wasting all the progress for the U-1 (uniformity minus one) number of peers in the hyperedge. e.g. (A, 97.2), (B, 97.25), (C, 97.4), but A, B, C are all at different points in hyperparameter and parameter space. Then there's no need to form a hyperedge. Weaker workers - possibly stuck in a local minima should train with stronger models to better utilize computation. Need a threshold for how far apart in performance peers must be in order to form a hyperedge.