Uncertainty-Weighted MCTS Playouts

[shindavid]

Used by KataGo, described here.

Basically, the MCTS player takes raw value/policy priors (P) and refines them through MCTS to get P'. The neural network is essentially tasked with predicting |P' - P|. In other words, its tasked with predicting how confident it is of its output. During MCTS, the backpropagated playout values are scaled by this confidence value.

David Wu claims that this, combined with #37, results in a 50 ELO improvement in KataGo.

Implement this, and validate its value through experiments.