[MODEL] support gemma4_unified

[HaozheZhang6]

Summary

gemma4_unified (multimodal Gemma 4, e.g. OpenYourMind/gemma-4-12B-it-abliterated-uncensored) fails with Unsupport model_type gemma4_unified, and failed to auto-detect module tree. It's a close sibling of the existing gemma4 composite model, but its text stack drops the per-layer input adapters, so it needs its own definition rather than reusing Gemma4ForConditionalGenerationGPTQ.

What Changed

• New Gemma4UnifiedForConditionalGenerationGPTQ definition over model.language_model.layers: per-projection q/k/v norms + dual pre/post feed-forward norms (same as gemma4), but no per_layer_input_gate / post_per_layer_input_norm / per_layer_projection — this variant has none. Its language_model also has no project_per_layer_inputs, which is why reusing the gemma4 class AttributeErrors during the calibration generate hook.
• prepare_layer_replay_kwargs refreshes the rope per layer: the model builds one (cos, sin) per attention layer_type and hands each decoder layer the tuple for its own type, so cached-layer replay has to regenerate it for the layer's layer_type (sliding vs full). Without it, replay reuses a stale tuple and hits a rope shape mismatch in apply_rotary_pos_emb.
• Registered gemma4_unified in MODEL_MAP and re-exported from definitions/__init__.py.

Vision/audio towers stay as base (non-quantized) modules, same as gemma4.

Tests

• I added a new simple/fast unit test for this change, or documented why that is not applicable.
• I ran the new targeted test locally before opening this PR.
• I ran any other directly relevant local tests.

tests/test_gemma4_unified_support.py (CPU-only): model_type → definition resolution, module tree excludes the per-layer-input paths, and the rope refresh returns the right tuple per layer_type.

$ pytest tests/test_gemma4_unified_support.py -v
tests/test_gemma4_unified_support.py::test_gemma4_unified_model_type_selects_definition PASSED
tests/test_gemma4_unified_support.py::test_gemma4_unified_module_tree_excludes_per_layer_input_paths PASSED
tests/test_gemma4_unified_support.py::test_gemma4_unified_replay_kwargs_refresh_position_embeddings_per_layer_type PASSED
3 passed in 3.93s

I also ran a full quantize → save → reload → generate on a tiny gemma4_unified (4-bit, group_size 128) on an A100: quantization completes across both the sliding and full attention layers, the saved model reloads with TritonV2Linear packed layers, and generation runs without error.

Review Requirements

• I personally reviewed every file in this diff.
• I checked that the code matches existing project structure, APIs, and conventions.
• I avoided unnecessary monkeypatching and used the project's normal extension points where possible.

Notes

The definition follows gemma4.py's rope-replay approach but is kept as a separate class/file since the per-layer-input machinery there doesn't apply here. I didn't have the full 12B checkpoint on hand, so the end-to-end run used a tiny synthetic gemma4_unified — happy to validate against a specific checkpoint if you'd like.

[Qubitium]

@HaozheZhang6 Thank you! LGTM.