Reinforcement Learning-Based Variable Speed Limits Control to Reduce Crash Risks near Traffic Oscillations on Freeways