Aspect-Guided Multi-Level Perturbation Analysis of Large Language Models in Automated Peer Review