- The size of second channel should be size[2] of output tensor,
- The Scalar should be {variance[0], variance[0], variance[0], variance[0]}
  for _variance.size() == 1 case.
Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>