From Seconds to Hours: Reviewing MultiModal Large Language Models on Comprehensive Long Video Understanding