Use tensor as the bottom implementation and support plain scalar.
Note: DO NOT support the backward for ReduceSumAll function.