Guiding Online Reinforcement Learning with Action-Free Offline Pretraining