All-in-one: Understanding and Generation in Multimodal Reasoning with the MAIA Benchmark