Add LEO mini/tiere-ImageNet using pretrained feature map

[IsakFalk]

LEO introduced a feature map trained on mini-imagenet. It's fairly common to benchmark on this "dataset" where the instances are replaced by the feature mapped instances (which is hardcoded into the pickle files they provided). It would be good to provide a way to integrate this variation of mini and tiered-ImageNet into the datasets that learn2learn provides.