deep copy not working with torch/auto-initialize

Bug Description

Im not quite sure if this is an issue with pyo3, but

import torch.nn as nn
encoder = nn.TransformerEncoderLayer(320, 4, dropout=0.0)
encoders = nn.TransformerEncoder(encoder, 3) <- uses deep copy 

works in python. heres another example for deep copy

>>> import torch.nn as nn
>>> encoder = nn.TransformerEncoderLayer(320, 4, dropout=0.0)
>>> import copy
>>> copy.deepcopy(encoder)
TransformerEncoderLayer(
  (self_attn): MultiheadAttention(
    (out_proj): NonDynamicallyQuantizableLinear(in_features=320, out_features=320, bias=True)
  )
  (linear1): Linear(in_features=320, out_features=2048, bias=True)
  (dropout): Dropout(p=0.0, inplace=False)
  (linear2): Linear(in_features=2048, out_features=320, bias=True)
  (norm1): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
  (norm2): LayerNorm((320,), eps=1e-05, elementwise_affine=True)
  (dropout1): Dropout(p=0.0, inplace=False)
  (dropout2): Dropout(p=0.0, inplace=False)
)

Steps to Reproduce

 Python::with_gil(|py| {
        // import torch.nn as nn
        let nn = py.import("torch.nn").unwrap();

        let kwargs = PyDict::new(py);
        kwargs.set_item("dropout", 0.0).unwrap();
        // encoder = nn.TransformerEncoderLayer(320, 4, dropout=0.0)
        let encoder = nn
            .getattr("TransformerEncoderLayer")
            .unwrap()
            .call((320, 4), Some(&kwargs))
            .unwrap();

        let cloned_module = copy
            .getattr("deepcopy")
            .unwrap()
            .call1((&encoder,))
            .unwrap();
    });

Backtrace

[1]    98639 segmentation fault  PYTHONPATH=venv/lib/python3.11/site-packages cargo r

Your operating system and version

MACOS 15.6

Your Python version (python --version)

Python 3.11.9

Your Rust version (rustc --version)

rustc 1.87.0 (17067e9ac 2025-05-09)

Your PyO3 version

0.25.1

How did you install python? Did you use a virtualenv?

venv

Additional Info

No response

Start by reproducing the macOS failure from the Python::with_gil block using PyO3 0.25.1, Python 3.11.9, and the shown torch.nn TransformerEncoderLayer example. Trace the deepcopy call and compare it with Python's successful copy.deepcopy behavior; done means the Rust embedding no longer segfaults and the cloned module is returned.