Random Field Augmentations for Self-Supervised Representation Learning