Exploring the Potential of Multimodal LLM with Knowledge-Intensive Multimodal ASR