[Request] nice-ifying build failures around hardware acceleration

[jamesbraza]

Build scripts support many types of hardware acceleration: https://github.com/abetlen/llama-cpp-python/blob/v0.2.11/README.md#installation-with-hardware-acceleration

# cuBLAS
CMAKE_ARGS="-DLLAMA_CUBLAS=on" pip install llama-cpp-python
# Metal
CMAKE_ARGS="-DLLAMA_METAL=on" pip install llama-cpp-python

If one is missing OS-level dependencies, the errors thrown can be somewhat arcane, to those not informed. The request is to nice-ify these messages, providing context and/or solutions.

For example, when installing on an AWS EC2 g4dn.4xlarge, without nvcc installed via nvidia-cuda-toolkit:

      CMake Error at /tmp/pip-build-env-xoh3dnhj/normal/lib/python3.11/site-packages/cmake/data/share/cmake-3.27/Modules/CMakeDetermineCUDACompiler.cmake:603 (message):
        Failed to detect a default CUDA architecture.

This can be nice-ified to be:

        Failed to detect a default CUDA architecture.
        cuBLAS backend failed: do you have nvcc installed?

Or when accidentally pasting in the wrong CMAKE_ARGS (here I mixed up Metal and cuBLAS):

      CMake Error at vendor/llama.cpp/CMakeLists.txt:174 (find_library):
        Could not find FOUNDATION_LIBRARY using the following names: Foundation

This can be nice-ified to be:

        Could not find FOUNDATION_LIBRARY using the following names: Foundation
        Metal backend failed: are you on a Mac?

In Python, Exceptions can be re-raised using from: raise Exception("Nice message.") from arcane_exc. I am not sure if this is something supported by CMakeLists.txt, but I think it would be really helpful to provide better failure messages for the hardware acceleration specifics.

For context, these errors were encountered in ggml-org/llama.cpp#3459.