Revisiting Cross-Lingual Summarization: A Corpus-based Study and A New Benchmark with Improved Annotation