Energy-Inspired Self-Supervised Pretraining for Vision Models