Fix libtorch_cuda_linalg builds (#73896)
When linking statically with cuda, link order matters
Link `magma` before `torch_cuda` to ensure all missing cublas symbols are resolved in either from `torch_cuda` or from `libcublas.a`
Add `culibos` dependency at the very end
Test plan: run `python -c "import torch; print(torch.rand(3,3,device='cuda').det())"` on installed manywheel
