LLMs as Adaptive Meta-Solvers: Strategy-Diverse RL for Industrial-Scale Optimization