STEP: State-Aware Task Estimation and Planning with Multi-Modal LLMs for Human-Robot Collaboration