auto-round-kernel installation method

[chensuyue]

This pull request introduces significant improvements to the AutoRound Kernel integration, including enabling and documenting the kernel backend, updating requirements and installation methods, and enhancing test coverage. It also refactors the ARK QLinear implementation for better device and dtype handling. The most important changes are summarized below:

AutoRound Kernel Integration and Documentation:

• Added a comprehensive README.md for AutoRound Kernel, detailing supported hardware, quantization configurations, versioning, and installation instructions.
• Introduced a new installation script install_kernel.py to automatically detect the PyTorch version and install the appropriate kernel version, and registered a new CLI command auto-round-kernel-install. [1] [2]

Backend Configuration and Requirements:

• Enabled and registered the auto_round_kernel, auto_round_kernel_zp, and auto_round_kernel_awq backends for CPU (previously commented out), and updated their PyTorch version requirements to torch>=2.8.0 for broader compatibility. [1] [2] [3]
• Removed the "kernel" extra from setup.py to simplify dependency management.

ARK QLinear Refactoring:

• Refactored ark/qlinear.py to instantiate ARK via auto_round_kernel.ARK(), improved dtype/device handling, unified bias and input dtype logic, and switched to using woqgemm for computation. [1] [2] [3]

Testing Improvements:

• Expanded test coverage in test_model.py to include CPU devices for all relevant test cases, and re-enabled tests for additional quantization configurations.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[chensuyue]

• Test on CPU must run with OMP_NUM_THREADS and numactl, e.g. OMP_NUM_THREADS=32 numactl -C "0-31" pytest -v test_ark/test_model.py.
• Test verified locally on GNR and BMG.
• Keep ark CI closed, because test failed on CI machine (EMR), debugging WIP.