Support lazy-initialized Linear modules in SVDQW4A4Linear.from_linear

[kj863257]

Motivation

Some integration paths may construct Linear modules lazily, with weight=None until state dict loading completes.

SVDQW4A4Linear.from_linear() currently assumes linear.weight is always available and directly accesses
linear.weight.dtype and linear.weight.device. In lazy-initialization scenarios, this can raise:

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

This PR adds safe fallback behavior for lazy-initialized Linear modules while preserving the existing behavior for
regular nn.Linear modules that already have materialized weights.

This change is part of a paired fix with the corresponding ComfyUI-nunchaku PR:

• ComfyUI-nunchaku: Support lazy-init Linear layers in Z-Image patching
• Link: <COMFYUI_NUNCHAKU_PR_URL>

Modifications

• Update SVDQW4A4Linear.from_linear() in nunchaku/models/linear.py.
• Keep the existing behavior unchanged when linear.weight is present.
• When linear.weight is None, resolve torch_dtype from the explicit argument if provided, otherwise fall back to
  linear.weight_comfy_model_dtype, and finally torch.bfloat16.
• Resolve device from the explicit argument if provided, otherwise fall back to weight.device, bias.device, and
  finally CPU.
• Determine bias presence from linear.bias instead of assuming a populated weight tensor.
• Add a lightweight unit test in tests/test_linear_lazy_init.py covering both lazy-initialized and regular Linear
  cases.
• Run pre-commit on the touched files.

Checklist

• Code is formatted using Pre-Commit hooks.
• Relevant unit tests are added in the tests directory following the guidance in Contribution Guide.
• Documentation and example scripts in examples are updated if necessary.
• Throughput/latency benchmarks and quality evaluations are included where applicable.
• 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 [Slack](https://join.slack.com/t/nunchaku/shared_invite/zt-3170agzoz-
  NgZzWaTrEj~n2KEV3Hpl5Q), Discord or [WeChat](https://huggingface.co/datasets/
  nunchaku-tech/cdn/blob/main/nunchaku/assets/wechat.jpg) to discuss your PR.

nunchaku-ai/ComfyUI-nunchaku#817