Fix gated Qwen W_Q analysis weights

[emerardd]

Description

Fixes #1652.

Qwen3.5 and Qwen3Next use a gated query projection whose rows are interleaved as query and gate values within each head. TransformerBridge.W_Q previously exposed the complete projection as if every row belonged to the query, so its final dimension was 2 * cfg.d_head. Weight-space analyses such as composition scores, QK analysis, and SVD therefore consumed query-plus-gate data under a query-only interface.

The adapters attempted to remove the gate rows in preprocess_weights(), but that helper matched Hugging Face-style keys while the real processing path supplies TransformerLens-style state-dict keys. More importantly, preprocessing the parameter would mutate the live projection and remove the gate needed by the model forward.

This change:

• makes AttentionBridge.W_Q return a query-only per-head view when gated_q_proj is enabled;
• leaves the underlying query-and-gate projection unchanged, preserving forward behavior and hook_q_gate;
• removes the unreachable gated-query preprocessing helper and tests built around synthetic key formats;
• adds download-free, config-only integration coverage for Qwen3.5 and Qwen3Next through build_bridge_from_module() and process_weights();
• checks exact query-row values, multi-layer W_Q shape, forward parity, gate-hook behavior, live-weight preservation, ordinary Qwen3 as a control, and the multimodal adapter's shared gated-query contract.

No new dependencies are required.

Type of change

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

Screenshots

Not applicable.

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Validation

• Full unit suite: 5260 passed, 55 skipped, 54 deselected, 10 xfailed
• Affected and adjacent Qwen/attention coverage: 146 passed
• New download-free integration regressions: 5 passed
• Formatting: pycln, isort, and Black pass
• Type checking: Success: no issues found in 431 source files
• git diff --check passes

[jlarson4]

Looks good, great resolution for this issue @emerardd!