feat(ml): add supervised loss with GANs with aligned datasets

[beniz]

This PR adds supervised loss to GAN training when dataset is aligned.

New options:

• --alg_cut_supervised_loss: empty, L1 or MSE
• --alg_cut_lambda_supervised: supervised loss weight, default to 1.0