MMDU: A Multi-T urn Multi-Image Dialog Understanding Benchmark and Instruction-T uning Dataset for L VLMs