Jigsaw Euclidean Embedding

Created by: lehr-fa

This implements the Euclidean embedding of the discrete permutation metric in the Jigsaw task. In consequence, the former classification problem becomes a regression problem, which is easier to be solved by the optimizer. Since multiprocessed data loading and multi-gpu training lead to multiple invocations of the training script, the embedding has to be computed beforehand and saved on disk to be consistent. If it was computed on-the-fly, the non-deterministic character of the embedding computation would lead to different embeddings on each data loader process and/or each GPU.