Multimodal Needle in a Haystack: Benchmarking Long-Context Capability of Multimodal Large Language Models