[MODEL] add diffusion_gemma quantization support

[ZX-ModelCloud]

Summary

• add native DiffusionGemma multimodal quantization support for diffusiongemma-26B-A4B-it
• quantize the encoder capture path and restore shared QuantLinear/dense aliases on the diffusion decoder
• support LazyTurtle checkpoint namespace remapping, packed MoE experts, attention-mask replay, and specialized generation config
• replace stale broad tied-weight rules with exact storage-backed mappings to prevent Transformers tie_weights_keys failures
• add image calibration/generation coverage and document the model in README
• remove the unused defuser_auto_detect_moe model flag

Dependency

• Requires defuser>=0.0.28 from feat: support DiffusionGemma expert defusion Defuser#67 for explicit DiffusionGemma expert defusion.

Validation

• conda run -n gp_311 pytest -q tests/test_diffusion_gemma_support.py — 17 passed