HuggingFaceModelInitializer and S3ModelInitializer share a mutable default ignore_patterns list across instances

### What happened?

Description

While reviewing kubeflow/trainer/types/types.py, I noticed that both HuggingFaceModelInitializer and S3ModelInitializer define ignore_patterns as:

ignore_patterns: list[str] | None = field(
default_factory=lambda: constants.INITIALIZER_DEFAULT_IGNORE_PATTERNS
)

At first glance this looks correct because default_factory is used. However, the factory returns the module-level mutable list directly instead of creating a new list for each instance.

INITIALIZER_DEFAULT_IGNORE_PATTERNS is defined as:

INITIALIZER_DEFAULT_IGNORE_PATTERNS = [
"*.msgpack",
"*.h5",
"*.bin",
"*.pt",
"*.pth",
]

Since the same list object is returned every time, all initializer instances share the same underlying ignore_patterns list.

Verified locally
from kubeflow.trainer.types.types import HuggingFaceModelInitializer

a = HuggingFaceModelInitializer(storage_uri="hf://user/model1")
b = HuggingFaceModelInitializer(storage_uri="hf://user/model2")

print(a.ignore_patterns is b.ignore_patterns)

Output:

True

Both instances reference the exact same list object.

To verify the impact:

print("Before:")
print("A:", a.ignore_patterns)
print("B:", b.ignore_patterns)

a.ignore_patterns.append("*.custom")

print("After:")
print("A:", a.ignore_patterns)
print("B:", b.ignore_patterns)

Output:

Before:
A: ['*.msgpack', '*.h5', '*.bin', '*.pt', '*.pth']
B: ['*.msgpack', '*.h5', '*.bin', '*.pt', '*.pth']

After:
A: ['*.msgpack', '*.h5', '*.bin', '*.pt', '*.pth', '*.custom']
B: ['*.msgpack', '*.h5', '*.bin', '*.pt', '*.pth', '*.custom']

Even though only a.ignore_patterns was modified, the change is immediately visible in b.ignore_patterns because both instances share the same mutable list.

Existing pattern in the codebase

The same file already uses the correct pattern for CustomTrainer.pip_index_urls:

pip_index_urls: list[str] = field(
default_factory=lambda: list(constants.DEFAULT_PIP_INDEX_URLS)
)

Creating a new list for every instance avoids sharing mutable state.

Expected behavior

Each HuggingFaceModelInitializer and S3ModelInitializer instance should receive its own independent copy of the default ignore_patterns list.

Suggested fix
ignore_patterns: list[str] | None = field(
default_factory=lambda: list(constants.INITIALIZER_DEFAULT_IGNORE_PATTERNS)
)

This matches the pattern already used elsewhere in the same file.

Environment

Repository: kubeflow/sdk
File: kubeflow/trainer/types/types.py
Affected classes:
HuggingFaceModelInitializer
S3ModelInitializer

I'd be happy to submit a PR if this behavior is considered a bug.

### What did you expect to happen?

Each `HuggingFaceModelInitializer` and `S3ModelInitializer` instance should receive its own independent copy of the default `ignore_patterns` list.

Mutating the `ignore_patterns` of one instance should not affect any other initializer instance.

For example, after creating two independent initializer objects, appending a new ignore pattern to one instance should leave the other instance unchanged.

### Environment

Kubernetes version:
```bash
$ kubectl version

```
Kubeflow Trainer version:
```bash
$ kubectl get pods -n kubeflow -l app.kubernetes.io/name=trainer -o jsonpath="{.items[*].spec.containers[*].image}"

```
Kubeflow Python SDK version:
```bash
$ pip show kubeflow

```

### Impacted by this bug?

Give it a 👍 We prioritize the issues with most 👍

Start in kubeflow/trainer/types/types.py at HuggingFaceModelInitializer and S3ModelInitializer, then compare their ignore_patterns definitions with CustomTrainer.pip_index_urls. Verify that separately created initializers no longer share mutable default state, and confirm that changing one instance leaves the other unchanged.

Written by the indexing model from the issue text.