Think Smart, Not Hard: Difficulty Adaptive Reasoning for Large Audio Language Models