fix test_evaluation and AutoTuner new issue

[xin3he]

Description

Please briefly describe your main changes, the motivation.

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #1227

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

This PR adjusts CUDA evaluation tests and centralizes a Triton Autotuner compatibility patch to address issues seen during evaluation and autotuning.

Changes:

• Set NCCL_ASYNC_ERROR_HANDLING=1 in CUDA evaluation tests that use vLLM.
• Remove per-module Autotuner patching from Triton dequant modules.
• Add a centralized Autotuner compatibility patch in auto_round_extension.triton package init.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
test/test_cuda/advanced/test_evaluation.py	Adds NCCL async error handling env var to stabilize vLLM-based eval runs.
auto_round_extension/triton/triton_utils_zp/dequant.py	Removes inline Autotuner patching side effect from the dequant module.
auto_round_extension/triton/triton_utils/dequant.py	Removes inline Autotuner patching side effect from the dequant module.
auto_round_extension/triton/init.py	Introduces centralized Autotuner patching on package import for older Triton versions.