AMDGPU tests are OOMing the runners

[kshyatt]

Thanks @luraess for surfacing this

We're being bad neighbors in the AMD GPU CI, our tests are tooooooo memory heavy and it's OOMing some of the Docker runners.

Seems it allocs about 60 GB if not wrong. Runners 1 has "only" 32 GB. Maybe that's one of the reason. Dunno tho as in CUDA tests memory usage spikes at 80 GB. But maybe the runner can accommodate this for CUDA

Quick fix: restrict to target gfx1101 in our runners.

Could be a good forcing function to make our tests more modular and do what the GPUArrays testsuite does.

[luraess]

I see you are using ParallelTestRunner.jl to orchestrate the tests. A first straight forward fix could be to just split the unique testfile for AMDGPU into a few smaller ones which would then be able to run in parallel on various runners an benefit from the implicit resource management built into the runner orchestration.