Fused QKV add node issue for GQA graph surgery (#1057)

### What does this PR do?

Type of change: Bug fix

There was a small issue where for models like qwen which have bias add
nodes, while fusing the q,k,v matmul and q,k,v add nodes , the fused qkv
bias add node was added to the graph before the fused qkv matmul node,
causing the removal script to assume that the fused matmul and the add
node were part of dead subgraph hence removing them. I just changed the
order in which they are added. Now there are no issues.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->

## Summary by CodeRabbit

* **Refactor**
* Optimized graph surgery operations for ONNX model processing by
adjusting node insertion timing during the multi-head to grouped-query
attention transformation, maintaining functional equivalence while
improving internal processing flow.

<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Signed-off-by: Hrishith Thadicherla <hthadicherla@nvidia.com>

Author: hthadicherla

modelopt/onnx/graph_surgery/gqa_replacement.py

@@ -707,7 +707,7 @@ def _find_node_by_pattern(pattern: str, op_type: str | None = None) -> onnx.Node
                 outputs=[qkv_add_output],
                 name=qkv_add_name,
             )
-            graph.node.append(qkv_add_node)
+            qkv_matmul_nodes.append(qkv_add_node)
 
             # Add value_info
             qkv_add_info = helper.make_tensor_value_info(