* Add a fake_quantize_op, which quantize an input tensor to a tensor with lower bits.