[Misc] Further polish gcn2conv.py (#6046)

Co-authored-by: Ubuntu <ubuntu@ip-172-31-28-63.ap-northeast-1.compute.internal>

python/dgl/nn/pytorch/conv/gcn2conv.py

@@ -254,13 +254,13 @@ class GCN2Conv(nn.Module):
 
             # initial residual connection to the first layer
             feat_0 = feat_0[: feat.size(0)] * self.alpha
+            feat_sum = feat + feat_0
 
             if self._project_initial_features:
-                feat_sum = feat + feat_0
-                rst = (1 - self.beta) * feat_sum + self.beta * feat_sum @ self.weight1
+                feat_proj_sum = feat_sum @ self.weight1
             else:
-                feat_sum = feat + feat_0
                 feat_proj_sum = feat @ self.weight1 + feat_0 @ self.weight2
+
             rst = (1 - self.beta) * feat_sum + self.beta * feat_proj_sum
 
             if self._bias: