IntentionVLA: Generalizable and Efficient Embodied Intention Reasoning for Human-Robot Interaction