Fix AWQ tests for GPTQModel migration

[jiqing-feng]

This PR fixes the AWQ test suite to align with the GPTQModel migration (following #41567 and #42776).

Changes

• Fix replace_with_awq_linear return value: The function now returns the model directly instead of a tuple (model, _), updated all call sites accordingly.
• Use BaseQuantLinear for type checking: Replaced specific AwqGEMMQuantLinear / AwqGEMVQuantLinear isinstance checks with the unified gptqmodel.nn_modules.qlinear.BaseQuantLinear.
• Remove explicit GEMM backend override in test setup: Let the model load with its default quantization config instead of forcing AwqBackend.GEMM.
• Fix save and load test: Load a fresh model for save/reload testing to avoid corrupted buffers caused by in-place transforms from prior generate() calls on the shared model instance.
• Update ground truth expected output: Added an additional valid expected output for bf16 inference.

@SunMarc @Qubitium

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: autoawq

[Qubitium]

LGTM! Allowing gptqmodel to auto load the awq kernel will also make future tests regress less since transformers no longer need to know about format=awq to correct kernel to env mappings.

[SunMarc]

Thanks !

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.