Policy Abstraction and Nash Refinement in Tree-Exploiting PSRO