Decoding Style: Efficient Fine-Tuning of LLMs for Image-Guided Outfit Recommendation with Preference