ConKeD++ - Improving descriptor learning for retinal image registration: A comprehensive study of contrastive losses.