Fragile Preferences: A Deep Dive Into Order Effects in Large Language Models