UNComp: Uncertainty-Aware Long-Context Compressor for Efficient Large Language Model Inference