add evaluation metrics for cross-modality networks and autoencoders

[borauyar]

Currently, the prediction performances are evaluated for regression/classification/survival tasks, but in the absense of any of these e.g. when using unsupervised mode or using cross-modality networks, we don't by default evaluate the reconstruction performance of the network of the target omic layers. We can compute the correlation scores between the known omics measurements and the predicted measurements row-wise or column-wise.