feat: support DiffusionGemma expert defusion

[ZX-ModelCloud]

Summary

• register DiffusionGemma fused experts explicitly
• defuse packed expert tensors into per-expert Linear modules with the linear-loop implementation
• add a minimal meta-model case covering both encoder and decoder expert blocks
• document diffusion_gemma in the supported-model matrix
• bump the package version to 0.0.28

Validation

• ruff check defuser/model_registry.py tests/test_meta_model_defusion.py
• conda run --no-capture-output -n gp_311 pytest -q tests/test_meta_model_defusion.py -k diffusion_gemma — 1 passed
• version, dependency, and README consistency check passed