Beyond Single-Step Updates: Reinforcement Learning of Heuristics with Limited-Horizon Search