Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation