MM-SpuBench: Towards Better Understanding of Spurious Biases in Multimodal LLMs