ONNX supports different input types in Pow, so type promotion is not needed.

The resulting graph is the following:

```py
ONNXProgram(
    model=
        <
            ir_version=9,
            opset_imports={'': 18, 'pkg.onnxscript.torch_lib.common': 1},
            producer_name='pytorch',
            producer_version='2.6.0a0+git59a1af5',
            domain=None,
            model_version=None,
        >
        graph(
            name=main_graph,
            inputs=(
                %"x"<FLOAT16,[3]>
            ),
            outputs=(
                %"pow_1"<FLOAT16,[3]>
            ),
        ) {
            0 |  # node_Constant_0
                 %"val_0"<?,?> ⬅️ ::Constant() {value=Tensor<FLOAT,[]>(array(2., dtype=float32), name=None)}
            1 |  # node_Pow_1
                 %"pow_1"<FLOAT16,[3]> ⬅️ ::Pow(%"x", %"val_0")
            return %"pow_1"<FLOAT16,[3]>
        }
...
    ,
    exported_program=
        ExportedProgram:
            class GraphModule(torch.nn.Module):
                def forward(self, x: "f16[3]"):
                     # File: /workspace/pytorch/test/onnx/exporter/test_small_models_e2e.py:53 in forward, code: return x**2.0
                    pow_1: "f16[3]" = torch.ops.aten.pow.Tensor_Scalar(x, 2.0);  x = None
                    return (pow_1,)

        Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='x'), target=None, persistent=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='pow_1'), target=None)])
        Range constraints: {}

)
```

Pull Request resolved: pytorch#139527
Approved by: https://github.com/titaiwangms