Using SupervisedTrainer for HoVerNet pipeline training requires a prepare_batch to prepare data differently than how it is being prepared in default_prepare_batch since HoVerNet target includes three different tensors.