Wrong IS-MCTS value

[WhiffleFish]

Current IS-MCTS implementation takes average of all samples yielding consistently low exploitability estimates.

This was done because root node may not be an action node for the exploiting player.

For example, in Kuhn Poker, the first node is a chance node, so $\max Q$ can't be calculated. Need to recursively search tree for first nodes corresponding to exploiting player and weight by reach probabilities.