One Representation to Rule Them All: Identifying Out-of-Support Examples in Few-shot Learning with Generic Representations