Fix GPTQ model loading in Transformers backend

[hmellor]

This PR:

• Adds the ability to ignore unexpected suffixes to the AutoWeightLoader
• In the Transformers backend, ignore the suffix ".bias" if the quant method is gptq, as is done in many of the non-auto weight loaders in models/
• Adds a small AWQ & GPTQ model to the Transformers backend quantization test

This method could be leveraged in other model implementations to reduce the need to manually implement weight loading. For now, this is left as a future task.

[gemini-code-assist]

Code Review

This pull request effectively addresses the GPTQ model loading issue in the Transformers backend by introducing a mechanism to ignore unexpected weight suffixes, specifically .bias for GPTQ models. The changes are well-implemented, extending the AutoWeightLoader to support suffix-based ignoring. Additionally, the inclusion of new quantization tests for AWQ and GPTQ models, along with refactoring the ROCm skip logic, significantly improves the test suite's coverage and robustness. Overall, this is a solid contribution that enhances model compatibility.

[Isotr0py]

LGTM, thanks for fixing!

[hmellor]

I've cancelled the build to prevent wasted CI time.

It's strange that https://buildkite.com/vllm/ci/builds/32718/steps/canvas?jid=01998a79-e1e5-495a-8ed4-750f01bf0250 failed, it did not fail locally for me.