Qwen model support

[cv-chaitali]

Summary

Adds support for Qwen-family decoder models in tuned-lens.

Changes

• Supports Qwen decoder models that expose base_model.norm and base_model.layers
• Adds test coverage for Qwen2, Qwen2-MoE, Qwen3, Qwen3-MoE, and Qwen3-Next
• Adds a small Qwen support note with a real Qwen/Qwen3-0.6B LogitLens result

Related issue: #143

[Copilot]

Pull request overview

This PR adds support for Qwen-family decoder models in tuned_lens.model_surgery by recognizing Qwen base-model structures that expose a decoder stack (layers) and final normalization (norm), unblocking LogitLens / PredictionTrajectory usage for Qwen models (as reported in issue #143).

Changes:

• Add a Qwen-family decoder detection helper and use it to route get_final_norm() and get_transformer_layers() to base_model.norm / base_model.layers.
• Extend the test model fixture to include tiny Qwen2/Qwen3 (dense + MoE + Next) configs instantiated via AutoModelForCausalLM.from_config.
• Add a short documentation note with an example LogitLens result for Qwen/Qwen3-0.6B.

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated no comments.

File	Description
tuned_lens/model_surgery.py	Adds Qwen decoder recognition and maps Qwen final norm/layers to base_model.norm and base_model.layers.
tests/conftest.py	Adds tiny Qwen config helpers and includes Qwen2/Qwen3 variants in the randomized small-model test fixture.
docs/qwen-support/README.md	Documents the motivating error/issue and summarizes the Qwen support changes and a sample result.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.