Learning Granularity-Aware Affordances from Human-Object Interaction for Tool-Based Functional Grasping in Dexterous Robotics