In the full model, we will see
        w(float) -- quant - *dequant         x -- quant --- *dequant --  *F.convTranspose2d --- *quant - dequant
        and the backend should be able to fuse the ops with `*` into a quantized conv2d
        N)
rW