COMODO: Cross-Modal Video-to-IMU Distillation for Efficient Egocentric Human Activity Recognition