Explore Activation Sparsity in Recurrent LLMs for Energy-Efficient Neuromorphic Computing