LMAct: A Benchmark for In-Context Imitation Learning with Long Multimodal Demonstrations