Unsupported node types fail to lower with “KeyError”

🐛 Describe the bug

When the QNN practitioner encounters a node that it does not have a visitor for, it fails with an internal error. Ideally, the partitioner should just not partition any unknown nodes without error, in order to allow them to fall back to the ExecuTorch portable library or another fallback backend.

Here are the operators that I've seen fail in this manner:
aten._adaptive_avg_pool3d.default, aten.adaptive_max_pool2d.default, aten.avg_pool3d.default, aten.div.Tensor_mode, aten.index_select.default, aten.log10.default, aten.log1p.default, aten.log2.default, aten.flip.default, aten.max_pool3d_with_indices.default, aten.median.default, aten.median.dim, aten.round.decimals, aten.le.Scalar, aten.trunc.default.
To clarify, QNN does not need implementations for these. We just want to be able to partially delegate models containing these operators.

I've included a small example model, along with the current stack trace. The current behavior is that the lowering process errors out. The expected behavior is that the partitioning process (via to_edge_transform_and_lower) completes without error and does not partition the unsupported nodes.

Example model:

class Model(torch.nn.Module):
    def __init__(
        self,
        output_size=(5, 5),
        return_indices=False,
    ):
        super().__init__()
        self.adaptive_maxpool = torch.nn.AdaptiveMaxPool2d(
            output_size=output_size,
            return_indices=return_indices,
        )

    def forward(self, x):
        return self.adaptive_maxpool(x)

Example stack trace:

2025-08-22T03:10:31.0328184Z     tester.to_edge_transform_and_lower(generate_etrecord=True)
2025-08-22T03:10:31.0329014Z   File "/pytorch/executorch/src/executorch/backends/test/harness/tester.py", line 220, in to_edge_transform_and_lower
...
2025-08-22T03:10:31.0348091Z   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py", line 226, in propose_partitions
2025-08-22T03:10:31.0348951Z     if self._is_node_supported(node) and node not in assignment:
2025-08-22T03:10:31.0349808Z   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py", line 87, in _is_node_supported
2025-08-22T03:10:31.0350642Z     return self.operator_support.is_node_supported(
2025-08-22T03:10:31.0351458Z   File "/pytorch/executorch/src/executorch/backends/qualcomm/partition/qnn_partitioner.py", line 100, in is_node_supported
2025-08-22T03:10:31.0352355Z     op_wrapper = self.node_visitors[node.target.__name__].define_node(
2025-08-22T03:10:31.0352836Z KeyError: 'aten.adaptive_max_pool2d.default'

With the QNN AOT bits set up and executorch installation run with QNN, you can also repro with

python -m executorch.backends.test.suite.runner -m operators --filter test_adaptive_maxpool2d_dtype_float32_qnn

Versions

Commit 335de46fa76866a76599def5e34296f0ee5f6106

cc @cccclai @cbilgin