[Bug Report] Fix dense component initialization for Pythia hook shapes

[jlarson4]

Summary

test_transformer_bridge_hook_shapes is parameterized over multiple models, but unconditionally skips Pythia at the top of the test body due to a known initialization issue with the dense component.

Affected test

• tests/integration/test_hook_shape_compatibility.py:127 has inline pytest.skip("Pythia architecture needs dense component initialization fix")

What the test verifies

That all hooks fire with the expected tensor shapes (matching HookedTransformer's contract) on bridge models.

Acceptance criteria

• The Pythia branch passes the same shape assertions as the other parameterized models.
• The inline skip is removed.

Where to start

• Pythia uses JointQKVPositionEmbeddingsAttentionBridge (rotary + joint QKV; same family as GPT-NeoX). Look at transformer_lens/model_bridge/supported_architectures/pythia.py and the dense o component construction in the adapter's _get_blocks().
• Run the test locally with the skip removed to see the actual failure mode first.

[Rishik00]

Hey @jlarson4 since I am working on the pythia adapter tests (citing #1302) I think I can also work on this and raise a PR fixing the issue. Does that work for you?

[jlarson4]

@Rishik00 certainly, feel free to take this on as well.