Uncertainty quantification in fine-tuned LLMs using LoRA ensembles