Boosting the Convergence of Reinforcement Learning-based Auto-pruning Using Historical Data