Support multiple labels in networks

[yd8534976]

Currently, we can only use (images, labels) pair to feed a model. It is not flexible for a complicated model.
Sometimes, we may face cases like (images, labels_1, labels_2..). We have to take it into account.

[yd8534976]

To support this feature, we have to modify blueoil/networks/base.py, blueoil/cmd/train.py, blueoil/datasets/dataset_iterator.py. But if we modify theses files directly, we would also have to change hundreds of files like #618 . If so, the PR will be too large to be reviewed. So we could add new files like blueoil/networks/base_multi.py, blueoil/cmd/train_multi.py, blueoil/datasets/dataset_iterator_multi.py instead of modifying them directly. If so, lines of code can be controlled and the PRs will be reviewable.

[yd8534976]

• create blueoil/datasets/dataset_iterator_multi.py
• create an example dataset class, it will be blueoil/datasets/widerface_multi.py
• create blueoil/networks/base_multi.py
• create blueoil/networks/object_detection/yolo_utils.py
• create blueoil/networks/object_detection/yolo_v2_multi.py
• create blueoil/cmd/train_multi.py
• create blueoil/configs/core/object_detection/yolo_v2_multi_widerface.py