Exploring Effective Distillation of Self-Supervised Speech Models for Automatic Speech Recognition