[ai-kit] Review ollama ebuild to better support NVIDIA GPU

In the last releases of ollama the CUDA driver uses the new dev-util/nvidia-cuda-toolkit package with different releases:

• CUDA 13
• CUDA 12
• CUDA 11 (to test)

Every release supplies support for different devices:

if(CUDA_VERSION VERSION_GREATER_EQUAL "13.0")
    # CUDA 13: 
    set(CMAKE_CUDA_ARCHITECTURES "75;80;86;89;90;100;120")
elseif(CUDA_VERSION VERSION_GREATER_EQUAL "12.0")
    # CUDA 12: 
    set(CMAKE_CUDA_ARCHITECTURES "61;70;75;80;86;89;90")
else()
    # CUDA 11: Max Ampere (86), include Maxwell (50)
    set(CMAKE_CUDA_ARCHITECTURES "50;52;61;70;75;80;86")
endif()

What is saw is that the native detection of CUDA GPU currently used in the ollama ebuild doesn't work correctly every time.
Only forcing CMAKE_CUDA_ARCHITECTURES with value 75 permit to have my GeForce RTX 2060 Mobile working correctly.

So what we need to do is:
a) permit to customize CUDA_ARCHITECTURES from users based on their GPUs. Keep them the control is better.
b) try to use the reported values on CMAKE_CUDA_ARCHITECTURES based on the available nvidia-cuda-toolkit package or just keep just point a

Inspect the ollama ebuild and its current CUDA GPU detection, then compare the requested CUDA 11, 12, and 13 architecture values with the available nvidia-cuda-toolkit releases. The work is done when users can customize CUDA_ARCHITECTURES and the ebuild handles the supported toolkit versions without relying on unreliable native detection.