LLM Evaluation as Tensor Completion: Low-Rank Efficiency and Uncertainty Quantification