Support lazy-init Linear layers in Z-Image patching

[kj863257]

Motivation

On recent ComfyUI Windows builds, some Linear modules can be lazily initialized and temporarily have weight=None
before model weights are fully loaded.

In the Z-Image loading path, ComfyUI-nunchaku patches attention and feed-forward modules before weight loading
completes. The current implementation assumes linear.weight is always available and directly accesses
linear.weight.dtype and linear.weight.device, which can raise:

AttributeError: 'NoneType' object has no attribute 'dtype'

This PR makes the Z-Image patching path compatible with lazy-initialized Linear modules while preserving the
existing behavior when real weights are already present.

This PR is paired with the corresponding nunchaku PR:

• nunchaku: Support lazy-initialized Linear modules in SVDQW4A4Linear.from_linear
• Link: <NUNCHAKU_PR_URL>

Modifications

• Add a helper in models/zimage.py to resolve dtype and device from a possibly lazy-initialized Linear module.
• Keep the existing behavior unchanged when linear.weight is available.
• Fall back to the provided torch_dtype, or weight_comfy_model_dtype, and finally CPU when linear.weight is None.
• Apply this logic to fuse_to_svdquant_linear() so the Z-Image feed-forward patching path no longer depends on
  linear.weight.dtype or linear.weight.device being immediately available.
• Run pre-commit on models/zimage.py.

No new workflow test case is added because this PR does not introduce a new workflow, model asset, custom node, or
output behavior. It only makes the existing Z-Image patching path tolerant to lazy-initialized Linear modules.

Checklist

• Code is formatted using Pre-Commit hooks (run pre-commit run --all-files).
• Relevant unit tests are added in the tests/workflows directory following the guidance in
  the Contribution Guide.
• Reference images are uploaded to PR comments and URLs are added to test_cases.json.
• Additional test data (if needed) is registered in test_data/inputs.yaml.
• Additional models (if needed) are registered in scripts/download_models.py
  and test_data/models.yaml.
• Additional custom nodes (if needed) are added to [.github/workflows/pr-test.yaml](../.github/workflows/pr-
  test.yaml).
• For reviewers: If you're only helping merge the main branch and haven't contributed code to this PR, please
  remove yourself as a co-author when merging.
• Please feel free to join our Discord or [WeChat](https://huggingface.co/
  datasets/nunchaku-tech/cdn/blob/main/nunchaku/assets/wechat.jpg) to discuss your PR.