Distribute wheels with cuBLAS support for all supported NVIDIA GPU architectures

[jllllll]

I recently discovered that llama-cpp-python can be compiled with cuBLAS support for all supported GPU architectures by setting the CUDAFLAGS environment variable to -arch=all. On Windows, I can use these commands in CMD:

set FORCE_CMAKE=1
set "CMAKE_ARGS=-DLLAMA_CUBLAS=on"
set "CUDAFLAGS=-arch=all -lcublas"
python -m pip install git+https://github.com/abetlen/llama-cpp-python

Note that, due to an issue with the current llama.cpp version in this repo, -lcublas has to be added as well in order to link the needed cuBLAS library. Setting the VERBOSE environment variable to 1 allows you to see the full output of the build process. Doing this, I can see that it is indeed building for all architectures as it shows a warning about the deprecated Kepler architectures.

The resulting wheel works on my own system, but that is to be expected. I have not been able to test if a wheel works on a different system.

This will greatly improve the user experience for text-generation-webui. Especially for Windows users due to eliminating the need for Visual Studio.

[jllllll]

Seems this method does not work with the latest llama.cpp due to them changing their CMakeLists.txt to use -arch=native.

[gjmulder]

Duplicate of #243