Wire CLIP vision-layer hook aliases for LLaVA adapters

[koriyoshi2041]

Problem

The LLaVA-family adapters declare attention and MLP hook aliases for CLIP vision encoder layers, but CLIPVisionEncoderLayerBridge does not register the corresponding submodules. The four aliases therefore cannot resolve.

Fix

Register the CLIP layer norms, q/k/v/output projections, and fc1/fc2 MLP projections with the same generalized bridge structure used by the SigLIP tower. The attention bridge receives the vision tower dimensions rather than the language-model dimensions. The now-obsolete strict xfails are removed for LLaVA, LLaVA-Next, and LLaVA-OneVision.

Test

• uv run pytest tests/unit/model_bridge/test_hook_alias_resolution.py tests/unit/model_bridge/supported_architectures/test_llava_adapter.py tests/unit/model_bridge/supported_architectures/test_llava_next_adapter.py tests/unit/model_bridge/supported_architectures/test_llava_onevision_adapter.py -q (192 passed, 1 skipped, 3 unrelated xfailed)
• make format
• uv run mypy transformer_lens/model_bridge/generalized_components/clip_vision_encoder.py
• git diff --check

Risk

The change is limited to CLIP vision-layer component registration and alias resolution. It does not change the wrapped Hugging Face forward path or claim end-to-end multimodal parity.

[jlarson4]

Hi @koriyoshi2041! Thanks for closing the CLIP half of this gap. A couple small items to address below before we can merge.

[jlarson4]

Looks good! Approved