[CI] Compile failure tests should be run on CPU-only runners

While experimenting with a new Blackwell CI job, I noticed the runtime of the `cub.test.device.radix_sort_decomposer_fail.lid_0` test:
```
Start 455: cub.test.device.radix_sort_decomposer_fail.lid_0
119/123 Test #455: cub.test.device.radix_sort_decomposer_fail.lid_0 .............. Passed 1258.18 sec
```
This seems excessive given other tests which take in the order of a few seconds (see full example [here](https://github.com/NVIDIA/cccl/actions/runs/19866932608/job/56935399566?pr=6841#step:4:529)).

@alliepiper reported that this test takes half as long on other runners (RTX A6000 or H100).

We should investigate whether their are some low hanging fruits to make the test faster.