Fix Gemma 4 MLP hook aliases

[koriyoshi2041]

Description

Fixes #1646.

Problem

The Gemma 4 adapter decomposes its dense MLP into gate_proj, up_proj, and down_proj, but wraps them in a plain GeneralizedComponent. The standard gated-MLP aliases (hook_pre, hook_pre_linear, and hook_post) are therefore missing even though the underlying module has the same projection structure as other supported gated MLPs.

Fix

Use the adapter's existing _gated_mlp() helper for Gemma 4. This preserves the same projection paths while installing the standard hook aliases. The unit test now checks both the bridge type and the exact alias mapping.

Test

• uv run pytest tests/unit/model_bridge/supported_architectures/test_gemma4_adapter.py -q (10 passed)
• uv run pytest tests/integration/model_bridge/test_gemma4_bridge.py -q (12 passed across PLE/KV-shared, dense K=V, and MoE fixtures)
• uv run mypy transformer_lens/model_bridge/supported_architectures/gemma4.py
• make check-format
• git diff --check

Risk

Low. The component retains the same mlp path and gate/up/down projection names. Gemma 4's delegated blocks still execute the Hugging Face forward path, and the three existing architecture fixtures retain logit parity. The change only adds the gated-MLP compatibility interface around the already-decomposed projections.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist

• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing focused tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

[jlarson4]

Looks great, thanks for putting this together @koriyoshi2041!