Measuring Physical-World Privacy Awareness of Large Language Models: An Evaluation Benchmark