(0.100.1) Set CUDA.jl compatibility to v5.9+

[ali-ramadhan]

CUDA.jl v5.9.0 includes JuliaGPU/CUDA.jl#2869 which resolves a bunch of issues here so I think we should set the [compat] entry for CUDA.jl to v5.9+.

Technically v5.8.4 also includes the fixes although there's only one commit difference between them so I went for v5.9.

Resolves #4047
Resolves #4694
Resolves #4663

[ali-ramadhan]

Environment fails to initialize on Buildkite. Looks like there's a package version conflict because CUDA.jl v5.9.0 wants PrettyTables.jl v3 but AMDGPU.jl v2.1.0 wants PrettyTables.jl v2.

The [compat] entry for AMDGPU.jl has already been updated to include PrettyTables.jl v3 (JuliaGPU/AMDGPU.jl#829) so I think we just need to wait for a new version of AMDGPU.jl to be tagged and released.

[navidcy]

Are we sure it actually resolve #4047, #4694, and #4663?

[tomchor]

Are we sure it actually resolve #4047, #4694, and #4663?

I can double-check, but it has for me since v5.8.4 in all the MWEs I tried.

I also think we should tag a patch release here since it solves these important bugs. What do people think?

[navidcy]

Sure, happy to tag a patch for this

[tomchor]

FYI just checked and this PR indeed solves all the above-mentioned issues (at least with the MWEs). I'll merge and tag a new version when tests pass.

[navidcy]

Note: I did run into a weird error with CUDA v5.9.0 when summing two CuArrays over a subset of their dimensions giving unexpectedly different results when repeated (despite the CuArrays content did not change).... That lead me to discovered that CUDA v5.8.4 and v5.9.0 suffer from JuliaGPU/CUDA.jl#2903 which was solved by JuliaGPU/CUDA.jl#2907 but is not yet included in a tagged release of CUDA.