Improve ABI compatibility for KTransformersOps extension

[zydxt]

This PR fixes ABI compatibility issues that caused undefined symbol errors when loading ktransformers.

Previously, KTransformersOps was compiled without explicitly setting the C++ ABI flag. This meant it might use a different ABI than installed PyTorch, leading to runtime errors like
KTransformersOps.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN3c106detail23torchInternalAssertFailEPKcS2_jS2_RKSs.

This change ensures that KTransformersOps is now compiled with the correct C++ ABI flag (-D_GLIBCXX_USE_CXX11_ABI=0 or 1) that matches PyTorch installation.

The core change involves updating setup.py to correctly apply the ABI flag to nvcc and mcc compilation steps, aligning with PyTorch's requirements.

The function get_compile_abi_args has been renamed to get_cmake_abi_args. While the original name might suggest it's only for CMake arguments, this function is now used to generate ABI-specific compiler flags for both CMake-based extensions and direct CUDAExtension/MUSAExtension compilations.

[Atream]

Thank you for your contribution!
But I think PyTorch Extension will automatically add the ABI parameter during compilation.
I will check it.