DensePose CSE with Cycle Losses

This release follows the paper Neverova et al, 2021 and adds CSE datasets with more annotations, better CSE animal models to the model zoo, losses to ensure cycle consistency for models and mesh alignment evaluator. In particular:

• Pixel to shape and shape to shape cycle consistency losses;
• Mesh alignment evaluator;
• Existing CSE datasets renamed to ds1_train and ds1_val;
• New CSE datasets ds2_train and ds2_val added;
• Better CSE animal models trained with the 16k schedule added to the model zoo.