Cyber-Attack Technique Classification Using Two-Stage Trained Large Language Models