Make elementwise_op supporting scalar input Y

[JiayiFeng]

solve #7568

[chengduoZH]

In this case, axis should be x->dims().size(), otherwise, it is wrong.
So I think this section of the code should be like this:

  int axis = ctx.Attr<int>("axis");
  axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);

  if (y_dims.size() == 1 && y_dims[0] == 1) {
    // y is a scalar
   axis = x_dims.size();
    auto extended_dims = framework::vectorize(x_dims);
    extended_dims.push_back(1);
    x_dims = framework::make_ddim(extended_dims);
  }

 

[JiayiFeng]

I don't quite agree with that. If the axis is inconsistent with the actual data shape, the code shall give an error instead of modifying axis in silence. Because we can never know which one is wrong, the axis or the data shape.

[chengduoZH]

LGTM

[gongweibao]

This PR is large. I think it's better to split elementwise_min_op elementwise_max_op elementwise with the support that Y is scalar into separate PRs.

Thanks @JiayiFeng, for the support that argument Y can be a scalar.

[JiayiFeng]

@gongweibao
I did. The PR of elementwise_max/min is #7538 .
This one depends on 7538. However, I can't merge 7538 yesterday because of the CI trouble. To avoid blocking my job, I merged 7538 to this PR on my own computer. So now this PR contains 7538.