Enable BUILD_SHARED_LIBS=ON on all Windows builds

[IsaacDynamo]

Enabled BUILD_SHARED_LIBS=ON on all Windows builds. With this change the release .zip files also contain llama.dll.

The release .dll allows you to use llama-cpp-python without compiling everything yourself. This is convenient for the more involved build targets such as cuBLAS.

llama-cpp-python will use the release .dll when the LLAMA_CPP_LIB env variable is set to the path of llama.dll

Tested the AVX2, OpenBLAS and cuBLAS .dll's with the following commands, and they work:

set LLAMA_CPP_LIB=C:\llama-release-dll-b1248-11eb9fb-bin-win-cublas-cu12.2.0-x64\llama.dll
set PATH=C:\llama-release-dll-b1248-11eb9fb-bin-win-cublas-cu12.2.0-x64;%PATH%
python.exe -m llama_cpp.server --model llama-2-7b-chat.Q5_K_M.gguf

[slaren]

Looks like this also reduces the size of the releases by a factor of 10, which is great.

[Green-Sky]

@IsaacDynamo did you check the cublas builds still work? they got suspiciously small :)

... also totally forgot i wanted to do this 😄

[slaren]

I tested the avx2 and cu12 executables and they work for me.