* discovery: only retry AMD GPUs CUDA and Vulkan don't crash on unsupported devices, so retry isn't necessary. This also refactors the code to shift the Library specific logic into the ml package. * review comments
