PENNI: Pruned Kernel Sharing for Efficient CNN Inference