Minimal Ranks, Maximum Confidence: Parameter-efficient Uncertainty Quantification for LoRA