Document solution of #526

[mofeing]

If the linker find a libcuda.so in a machine that doesn't have NVIDIA GPUs, Reactant will falsely believe that a CUDA is available and it will download the wrong artifact. The solution to #526 is to add a preference for Reactant_jll to use the gpu=none artifact variant.

This should be documented in the docs, given that this situation can happen to many HPC cluster users.

cc @giordano @Todorbsc

[giordano]

Any idea where this could live? Documentation seems to be a bit random at the moment.

[mofeing]

IMO it should be in a "Troubleshooting" page, but I don't know if that page should be in the general manual or in the docs about the internals.

[giordano]

Related, @wsmoses suggested to document how to debug the version of Reactant installed, to check whether CUDA is supported and the version picked up.

Leaving here some information for reference:

julia> using Libdl, Reactant_jll

julia> ENV["JULIA_DEBUG"] = "Reactant_jll"
"Reactant_jll"

julia> Reactant_jll.host_platform
Linux x86_64 {cuda_version=12.1, cxxstring_abi=cxx11, gpu=cuda, julia_version=1.11.3, libc=glibc, libgfortran_version=5.0.0, libstdcxx_version=3.4.30, mode=opt}

julia> Reactant_jll.cuDriverGetVersion(dlopen("libcuda.so"))
┌ Debug: Detected CUDA Driver version 12.2.0
└ @ Reactant_jll ~/.julia/packages/Reactant_jll/TDalS/.pkg/platform_augmentation.jl:60
v"12.2.0"