mobilebert fails returns bad results on CoreML

🐛 Describe the bug

Result is nan on FP16

sh .ci/scripts/test_model.sh mobilebert cmake coreml-pybind-float16

And result has high MSE with eager on FP32:

sh .ci/scripts/test_model.sh mobilebert cmake coreml-pybind-float32

Versions

Collecting environment information...
PyTorch version: 2.9.0.dev20250725
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: macOS 15.5 (arm64)
GCC version: Could not collect
Clang version: 17.0.0 (clang-1700.0.13.5)
CMake version: version 3.31.6
Libc version: N/A

Python version: 3.10.18 (main, Jun 5 2025, 08:37:47) [Clang 14.0.6 ] (64-bit runtime)
Python platform: macOS-15.5-arm64-arm-64bit
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
Is XPU available: False
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Apple M1 Pro

Versions of relevant libraries:
[pip3] executorch==0.8.0a0+8e6c8ff
[pip3] mypy_extensions==1.1.0
[pip3] numpy==2.2.6
[pip3] optimum-executorch==0.0.0.dev0
[pip3] pytorch_tokenizers==0.1.0
[pip3] torch==2.9.0.dev20250725
[pip3] torchao==0.13.0+git2eb4f9762
[pip3] torchaudio==2.8.0.dev20250725
[pip3] torchdata==0.11.0
[pip3] torchsr==1.0.4
[pip3] torchtune==0.6.1
[pip3] torchvision==0.24.0.dev20250725
[conda] executorch 0.8.0a0+8e6c8ff pypi_0 pypi
[conda] numpy 2.2.6 pypi_0 pypi
[conda] optimum-executorch 0.0.0.dev0 pypi_0 pypi
[conda] pytorch-tokenizers 0.1.0 pypi_0 pypi
[conda] torch 2.9.0.dev20250725 pypi_0 pypi
[conda] torchao 0.13.0+git2eb4f9762 pypi_0 pypi
[conda] torchaudio 2.8.0.dev20250725 pypi_0 pypi
[conda] torchdata 0.11.0 pypi_0 pypi
[conda] torchsr 1.0.4 pypi_0 pypi
[conda] torchtune 0.6.1 pypi_0 pypi
[conda] torchvision 0.24.0.dev20250725 pypi_0 pypi

cc @kimishpatel @YifanShenSZ @cymbalrush

Start by running the two .ci/scripts/test_model.sh commands for mobilebert and compare the CoreML FP16 and FP32 results with eager execution. Trace the mobilebert CoreML conversion or runtime path implicated by those commands. Done means FP16 no longer returns NaN and FP32 no longer has high MSE against eager.