Add example: pic breeder with contrastive loss on an incrementally growing dataset

[lukstafi]

The dataset consisting of (positive, negative) output pairs, and a Human Feedback UI to grow it via inference on the model while the model is training.