Pretraining with Random Noise for Fast and Robust Learning without Weight Transport