Neuromorphic Principles for Efficient Large Language Models on Intel Loihi 2