ProtoGAN: Towards high diversity and fidelity image synthesis under limited data