[Graph Optimizer] LayerNorm fusion incorrectly matches ReduceMean with keepdims=0, causing 'Node input not found' error

### Describe the issue

When loading a simple ONNX model that computes a normalized output using a sequence of operations (`ReduceMean`, `Sub`, `Pow`, `ReduceMean`, `Add`, `Sqrt`, `Div`, `Mul`), ONNX Runtime fails with an `InvalidArgument` error:

```
[ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Invalid model. Node input 'nm' is not a graph input, initializer, or output of a previous node.
```
Related onnx issue: [https://github.com/onnx/onnx/issues/8204](https://github.com/onnx/onnx/issues/8204)

### To reproduce

```python
from onnx import helper, TensorProto
import onnxruntime as ort

eps_t = helper.make_tensor("eps_v", TensorProto.FLOAT, [], [1.0e-5])
two_t = helper.make_tensor("two_v", TensorProto.FLOAT, [], [2.0])

nodes = [
helper.make_node('Constant', inputs=[], outputs=['eps'], value=eps_t),
helper.make_node('Constant', inputs=[], outputs=['two'], value=two_t),
helper.make_node('ReduceMean', inputs=['x'], outputs=['mn'], axes=[-1], keepdims=0), # keepdims=0
helper.make_node('Sub', inputs=['x', 'mn'], outputs=['ct']),
helper.make_node('Pow', inputs=['ct', 'two'], outputs=['sq']),
helper.make_node('ReduceMean', inputs=['sq'], outputs=['vr'], axes=[-1], keepdims=0), # keepdims=0
helper.make_node('Add', inputs=['vr', 'eps'], outputs=['ve']),
helper.make_node('Sqrt', inputs=['ve'], outputs=['sd_raw']),
helper.make_node('Identity', inputs=['sd_raw'], outputs=['sd']),
helper.make_node('Div', inputs=['ct', 'sd'], outputs=['nm']),
helper.make_node('Mul', inputs=['nm', 'z'], outputs=['out']),
]

graph = helper.make_graph(nodes, 'graph_0', [
helper.make_tensor_value_info('x', TensorProto.FLOAT, [2]),
helper.make_tensor_value_info('z', TensorProto.FLOAT, [2, 1]),
], [helper.make_tensor_value_info('out', TensorProto.FLOAT, [2, 2])])

model = helper.make_model(graph, opset_imports=[helper.make_opsetid('', 11)])
sess = ort.InferenceSession(model.SerializeToString())
```

### Urgency

Not urgent. This is a bug found by fuzzing testing.

### Platform

Linux

### OS Version

Ubuntu 24.04

### ONNX Runtime Installation

Released Package

### ONNX Runtime Version or Commit ID

1.20.1

### ONNX Runtime API

Python

### Architecture

X64

### Execution Provider

Default CPU

### Execution Provider Library Version

_No response_

Start by running the Python InferenceSession reproducer against the graph containing ReduceMean nodes with keepdims=0. Inspect the LayerNorm fusion matching path and verify that this pattern is not treated as a valid fusion when it leaves an unresolved input. Done means the model loads without the InvalidArgument 'Node input not found' error.

Written by the indexing model from the issue text.