Bump llama-cpp-python to 0.2.23 (including Linux ROCm and MacOS >= 12)

[oobabooga]

• Point the wheels back to https://github.com/jllllll/llama-cpp-python-cuBLAS-wheels
• Add the updated ROCm wheels for Linux.
• Add the updated MacOS wheels. A problem is happening with GitHub Actions for the MacOS 11 wheels, so the update only applies to MacOS 12 and 13 for now.

Full cretits to @jllllll for the update and for the GitHub Actions workflows that compile llama.cpp for hundreds of combinations of operating systems, GPU vendors, and Python versions.