Simplify backward when inserting a sum operator to accumulate all duplicated variables

[Xreki]

Since sum_op is used instead of add_op, and it supports to accumulate a vector of variables at one time, we do not need to insert multiple sum_op to add those inputs one by one. Just add all inputs in one sum_op.

I make a simple test based on interp_op. See diff of interp_op:

$ git diff interp_op.cc
diff --git a/paddle/operators/interp_op.cc b/paddle/operators/interp_op.cc
index d02b01c..08302b7 100644
--- a/paddle/operators/interp_op.cc
+++ b/paddle/operators/interp_op.cc
@@ -54,6 +54,10 @@ class InterpOp : public NetOp {
     // Out = MulOut + Y = (X - Y) * W + Y = X * W + Y * (1 - W)
     AppendOp(framework::OpRegistry::CreateOp("elementwise_add",
                                              {{"X", {mul_out}}, {"Y", {y}}},
+                                             {{"Out", {Output("SubOut")}}}, {}));
+
+    AppendOp(framework::OpRegistry::CreateOp("elementwise_add",
+                                             {{"X", {sub_out}}, {"Y", {y}}},
                                              {{"Out", {Output("Out")}}}, {}));
 
     CompleteAddOp(false);

The DebugString of interp_op:

147: Op(interp), inputs:{W[W], X[X], Y[Y]}, outputs:{MulOut[MulOut], Out[Out], SubOut[SubOut]}.
147: Op(elementwise_sub), inputs:{X[X], Y[Y]}, outputs:{Out[SubOut]}.
147: Op(elementwise_mul), inputs:{X[SubOut], Y[W]}, outputs:{Out[MulOut]}.
147: Op(elementwise_add), inputs:{X[MulOut], Y[Y]}, outputs:{Out[SubOut]}.
147: Op(elementwise_add), inputs:{X[SubOut], Y[Y]}, outputs:{Out[Out]}.

The DebugString of interp_grad_op in previous implementation:

147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_add_grad), inputs:{Out[Out], Out@GRAD[Out@GRAD], X[SubOut], Y[Y]}, outputs:{X@GRAD[SubOut@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_add_grad), inputs:{Out[SubOut], Out@GRAD[SubOut@GRAD], X[MulOut], Y[Y]}, outputs:{X@GRAD[MulOut@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_mul_grad), inputs:{Out[MulOut], Out@GRAD[MulOut@GRAD], X[SubOut], Y[W]}, outputs:{X@GRAD[SubOut@GRAD], Y@GRAD[W@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_sub_grad), inputs:{Out[SubOut], Out@GRAD[SubOut@GRAD], X[X], Y[Y]}, outputs:{X@GRAD[X@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(sum), inputs:{X[Y@GRAD@RENAME@0@0, Y@GRAD@RENAME@0@1]}, outputs:{Out[Y@GRAD@SHARED@0]}.
147: Op(sum), inputs:{X[Y@GRAD@RENAME@0@1, Y@GRAD@SHARED@0]}, outputs:{Out[Y@GRAD]}.
147: Op(sum), inputs:{X[SubOut@GRAD@RENAME@0@0, SubOut@GRAD@RENAME@0@1]}, outputs:{Out[SubOut@GRAD]}.

The DebugString of interp_grad_op in this PR:

147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_add_grad), inputs:{Out[Out], Out@GRAD[Out@GRAD], X[SubOut], Y[Y]}, outputs:{X@GRAD[SubOut@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_add_grad), inputs:{Out[SubOut], Out@GRAD[SubOut@GRAD], X[MulOut], Y[Y]}, outputs:{X@GRAD[MulOut@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_mul_grad), inputs:{Out[MulOut], Out@GRAD[MulOut@GRAD], X[SubOut], Y[W]}, outputs:{X@GRAD[SubOut@GRAD], Y@GRAD[W@GRAD]}.
147: Op(plain_net), inputs:{}, outputs:{}.
147: Op(elementwise_sub_grad), inputs:{Out[SubOut], Out@GRAD[SubOut@GRAD], X[X], Y[Y]}, outputs:{X@GRAD[X@GRAD], Y@GRAD[Y@GRAD]}.
147: Op(sum), inputs:{X[Y@GRAD@RENAME@0@0, Y@GRAD@RENAME@0@1, Y@GRAD@RENAME@0@2]}, outputs:{Out[Y@GRAD]}.
147: Op(sum), inputs:{X[SubOut@GRAD@RENAME@0@0, SubOut@GRAD@RENAME@0@1]}, outputs:{Out[SubOut@GRAD]}.

[luotao1]

LGTM. much simpler!