AdaFlow: Imitation Learning with Variance-Adaptive Flow-Based Policies