[Inductor][XPU] control_deps node used before definition in test_barrier_deps_exclude_nodes_defined_after_sync

### 🐛 Describe the bug

`TestUserStreamCompile::test_barrier_deps_exclude_nodes_defined_after_sync` in `test/inductor/test_user_streams.py` fails during AOT graph capture:

```
RuntimeError: Argument '_tensor_constant0' of Node 'control_deps' was used before
it has been defined! Please check that Nodes in the graph are topologically ordered
```

The `control_deps` higher-order op references `%_tensor_constant0`, but the `get_attr[target=_tensor_constant0]` node appears later in the graph, so `fw_module.graph.eliminate_dead_code()` -> `lint()` raises. Graph excerpt:

```
%control_deps : call_function[target=torch.ops.higher_order.control_deps](args = ((%mm, %_tensor_constant0), %subgraph_synchronize_stream, %mm, %_tensor_constant0))
...
%_tensor_constant0 : get_attr[target=_tensor_constant0]
```

### Repro

```
python test/inductor/test_user_streams.py TestUserStreamCompile.test_barrier_deps_exclude_nodes_defined_after_sync
```

### Versions

Device: XPU (`USE_XPU`), Triton Intel backend.

Start with test/inductor/test_user_streams.py and run TestUserStreamCompile.test_barrier_deps_exclude_nodes_defined_after_sync using the provided command. Inspect the AOT graph-capture path around eliminate_dead_code() and lint(); done means the XPU test passes without a control_deps node referencing _tensor_constant0 before its get_attr definition.

Written by the indexing model from the issue text.