[architectures] modified the original DR-GAN to be as close as possible from MSU (projection at the beginning of decoder)

[trainer] DR-GAN: function to get fixed data and loading models, fixed the images saved at each epoch (embedding the input image)

[scripts] fixed stuff related to model selection (and add the latent dim param needed for the original architecture)

[script] change the name of the script displaying training images, fixed default value for start epoch

[scripts] fixed the script to train using Multi-PIE and CASIA: corrected # of ids and start_index in for ID labels, use my own concatenation class

[datasets] made sure that id labels are native python int, since otherwise it could cause problem in the concatenation of datasets