BRIDGE: Benchmarking Large Language Models for Understanding Real-world Clinical Practice Text