Cross entropy stopping condition for training.

[Tilps]

Rather than using a fixed 800 nodes, or using pruning to do early stopping (which based on its selection point favors certain distribution shapes), I wonder if we could use a distribution cross entropy approach.
Previously some analysis was done that showed about 800 nodes was the peak cross entropy delta rate and afterwards has diminishing returns - but that was averaged over many games. If we sample the visit distribution every x nodes, we can calculate an average cross entropy per node change for the last x nodes. But specific for the current position, which may have a longer period of information gain compared to average.

Then we can set a minimum and maximum node visit range, and in between we stop if the average cross entropy per node change drops below a configurable value y.
Would need to do some simulations to find the trade off between speed and what value of y we can go down to, but maybe once we have draw agreements in place the additional saved time per game can be spent a bit here.

[Naphthalin]

#721 was merged a while ago, this issue can be closed I guess :)

[Tilps]

yeap, done.