CoreML int32 mm with fixed arg fails to load at runtime

🐛 Describe the bug

The following model fails to load at runtime with the following error:

[ETCoreMLModelManager.mm:528] Cache Miss: Model with identifier=executorch_c6fd42cf-7a82-4833-969f-e39c113b0890_all was not found in the models cache.
[ETCoreMLModelLoader.mm:69] [Core ML] Failed to load model from compiled asset with identifier = executorch_c6fd42cf-7a82-4833-969f-e39c113b0890_all Failed to build the model execution plan using a model architecture file '[/Users/gjcomer/Library/Caches/executorchcoreml/models/execu](http://localhost:8888/Users/gjcomer/Library/Caches/executorchcoreml/models/execu)$
[backend_delegate.mm:288] [Core ML] Model init failed The file “model.mlmodelc” couldn’t be opened because there is no such file.
[coreml_backend_delegate.mm:193] CoreMLBackend: Failed to init the model.
[method.cpp:113] Init failed for backend CoreMLBackend: 0x23

Repro:

import torch

from executorch.backends.apple.coreml.partition import CoreMLPartitioner
from executorch.exir import to_edge_transform_and_lower
from executorch.extension.pybindings.portable_lib import _load_for_executorch_from_buffer

class Model(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.weight = torch.randint(0, 100, (8, 8)).to(torch.int32)
    def forward(self, x):
        return torch.mm(x, self.weight)

model = Model()
inputs = (
    torch.randn(8, 8).to(torch.int32),
)

eager_outputs = model(*inputs)
print(f"Eager: {eager_outputs.shape} {eager_outputs}")

lowered = to_edge_transform_and_lower(
    torch.export.export(model, inputs),
    partitioner=[CoreMLPartitioner()],
).to_executorch()

et_model = _load_for_executorch_from_buffer(lowered.buffer)
et_outputs = et_model([*inputs])[0]

et_outputs - eager_outputs

Versions

coremltools version 8.3
executorch commit https://github.com/pytorch/executorch/commit/67b6009d6b3b67eee775c8ed2fe30eae6e0bb65c (Jun 14)

cc @kimishpatel @YifanShenSZ @cymbalrush @metascroy

Start by reproducing the failure with CoreMLPartitioner and the provided int32 torch.mm model. Trace the reported load path through ETCoreMLModelManager.mm, ETCoreMLModelLoader.mm, backend_delegate.mm, and coreml_backend_delegate.mm. Done means the lowered model loads at runtime and its output can be compared with eager_outputs.