SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehension