Efficient Monte Carlo Tree Search via On-the-Fly State-Conditioned Action Abstraction