[Fix] Load memdecode router checkpoints with weights_only

[zcxGGmu]

Motivation

Issue #3255 reports unsafe torch.load(...) usage for .pt checkpoint loading. The original load_weight_ckpt helper no longer exists on current main, but the memdecode adaptive-router path still loads user-provided .pt router checkpoints without explicitly setting weights_only=True.

Refs #3255

Modification

• Load memdecode router checkpoint metadata with torch.load(..., weights_only=True).
• Load memdecode router state dicts with torch.load(..., weights_only=True).
• Add regression coverage to ensure both router checkpoint load paths pass weights_only=True explicitly.

BC-breaking (Optional)

No. Supported router checkpoints using tensors and primitive dict metadata continue to load; unsafe pickle globals are rejected earlier by PyTorch's weights-only loader.

Use cases (Optional)

N/A

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
   • python -m pre_commit run --files lmdeploy/pytorch/memdecode/fusion.py tests/pytorch/memdecode/test_fusion.py
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
   • python -m pytest tests/pytorch/memdecode/test_fusion.py -q
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
   • N/A
4. The documentation has been modified accordingly, like docstring or example tutorials.
   • N/A; this is an internal checkpoint-loading hardening change.

[lvhan028]

I don't think we need TC for this. The change is a single parameter addition to torch.load; there's no business logic to regress.