[Sparse] Reduce peak memory of SpSpAdd. (#5143)

* [Sparse] Reduce peak memory usage of SpSpAdd

* Update

* Update

dgl_sparse/src/elemenwise_op.cc

@@ -22,9 +22,15 @@ c10::intrusive_ptr<SparseMatrix> SpSpAdd(
     const c10::intrusive_ptr<SparseMatrix>& A,
     const c10::intrusive_ptr<SparseMatrix>& B) {
   ElementwiseOpSanityCheck(A, B);
+  torch::Tensor sum;
+  {
+    // TODO(#5145) This is a workaround to reduce peak memory usage. It is no
+    // longer needed after we address #5145.
     auto torch_A = COOToTorchCOO(A->COOPtr(), A->value());
     auto torch_B = COOToTorchCOO(B->COOPtr(), B->value());
-  auto sum = (torch_A + torch_B).coalesce();
+    sum = torch_A + torch_B;
+  }
+  sum = sum.coalesce();
   auto indices = sum.indices();
   auto row = indices[0];
   auto col = indices[1];