Run CI on gpu #824

There are many ways to do that but I'd like one that is:

• Simple
• Reliable (need the gpu to be available 99+% of the time we need it)
• Future-proof (not some service that will be free for 3 months and then we have to pay)
• Free or cheap (up to maybe 5 usd / month, because we will only use the gpu for ~3 hours per month)
• Recent enough (need to support recent versions of torch, to support triton, etc, which my gtx 1080 will stop supporting soon)
• For now we should focus on cuda gpus, since they're the most used devices of torch users I think. In the future we could add support for rocm gpu testing, tpu, etc.

If we have a reliable solution, we could run the GPU check similarly as the CPU checks (on PR, daily or manually). If not, we could make it only run manually (workflow_dispatch).