[Arm backend] can't convert yolo models with aot_arm_compiler

🐛 Describe the bug

I'm using the latest release version v0.6.And I want to convert yolo models with arm backend.
I modified the examples as bellow, and convert the model with the command:
python3 -m examples.arm.aot_arm_compiler --model_name="yolo" --delegate --quantize --so_library="$(find arm_test -name libquantized_ops_aot_lib.so)"

    diff --git a/examples/arm/aot_arm_compiler.py b/examples/arm/aot_arm_compiler.py
    index 5fb12342a..a9f216cc2 100644
    --- a/examples/arm/aot_arm_compiler.py
    +++ b/examples/arm/aot_arm_compiler.py
    @@ -29,6 +29,7 @@ from executorch.backends.arm.quantizer.arm_quantizer import (
         get_symmetric_quantization_config,
         TOSAQuantizer,
     )
    +from ultralytics import YOLO
     from executorch.backends.arm.tosa_partitioner import TOSAPartitioner
     from executorch.backends.arm.tosa_specification import TosaSpecification
    
    @@ -82,6 +83,10 @@ def get_model_and_inputs_from_name(
             model = models[model_name]()
             if example_inputs is None:
                 example_inputs = models[model_name].example_input
    +    elif model_name == "yolo":
    +        model = YOLO("yolov5s.pt").model
    +        example_inputs = (torch.randn(1, 3, 640, 640),)
     
         elif model_name in MODEL_NAME_TO_MODEL.keys():
             logging.warning(
    @@ -648,7 +653,7 @@ def to_edge_TOSA_delegate(
             )
             model_int8 = model
             # Wrap quantized model back into an exported_program
    -        exported_program = torch.export.export_for_training(model, example_inputs)
    +        exported_program = torch.export.export_for_training(model, example_inputs, strict=False)
    

INFO:executorch.backends.arm.tosa_partitioner:TOSAPartitioner::partition
INFO:executorch.backends.arm.tosa_partitioner:Partitioning for EthosUBackend: TOSA-0.80+BI+u55
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/home/bamboo/latest/executorch/examples/arm/aot_arm_compiler.py", line 696, in <module>
    model_int8, edge = to_edge_TOSA_delegate(exported_program, args, model)
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/examples/arm/aot_arm_compiler.py", line 668, in to_edge_TOSA_delegate
    edge = to_edge_transform_and_lower(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/program/_program.py", line 107, in wrapper
    return func(self, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/program/_program.py", line 1209, in to_edge_transform_and_lower
    edge_manager = edge_manager.to_backend({name: curr_partitioner})
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/program/_program.py", line 107, in wrapper
    return func(self, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/program/_program.py", line 1471, in to_backend
    new_edge_programs[name] = to_backend(program, partitioner[name])
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/functools.py", line 909, in wrapper
    return dispatch(args[0].__class__)(*args, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/backend/backend_api.py", line 377, in _
    partitioner_result = partitioner_instance(fake_edge_program)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/exir/backend/partitioner.py", line 66, in __call__
    return self.partition(exported_program)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/backends/arm/tosa_partitioner.py", line 90, in partition
    partition_list = capability_partitioner.propose_partitions()
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/torch/fx/passes/infra/partitioner.py", line 221, in propose_partitions
    if self.__is_node_supported(node) and node not in assignment:
       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/torch/fx/passes/infra/partitioner.py", line 84, in __is_node_supported
    return self.operator_support.is_node_supported(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/torch/fx/passes/operator_support.py", line 153, in is_node_supported
    return is_node_supported(submodules, node)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/torch/fx/passes/operator_support.py", line 166, in _chain
    return all(x.is_node_supported(submods, node) for x in op_support)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/torch/fx/passes/operator_support.py", line 166, in <genexpr>
    return all(x.is_node_supported(submods, node) for x in op_support)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/backends/arm/operator_support/tosa_supported_operators.py", line 471, in is_node_supported
    tensor = get_first_fake_tensor(input_node)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/bamboo/latest/executorch/.venv/lib/python3.12/site-packages/executorch/backends/arm/_passes/arm_pass_utils.py", line 158, in get_first_fake_tensor
    assert isinstance(
           ^^^^^^^^^^^
AssertionError: Found zuf38 in meta["val"] of _local_scalar_dense_2, expected to find FakeTensor.

Versions

commit e67ef3b149688b4085605495690d0d5c6420b45a (HEAD -> main, tag: v0.6.0-rc9, tag: v0.6.0)

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218

Reproduce the command from examples/arm/aot_arm_compiler.py with the YOLO model and inspect the TOSA partitioning path. Start at tosa_supported_operators.py and _passes/arm_pass_utils.py, focusing on the _local_scalar_dense_2 metadata assertion. Done means the reported YOLO conversion no longer fails at this assertion and the ARM compilation flow completes.