AMR does not work on GPUs

[maximumcats]

When using inputs.128 with amr.max_level = 1 on summitdev, bad data is generated and the timesteps get really small. This does not occur with inputs.64, and does not occur in debug mode.

[maximumcats]

AMR now works without crashing and gives answers that are at least in the right ballpark, but the answers are still a bit different than on the CPU.

[maximumcats]

It is still the case that with AMR enabled, the GPU version gets different results than the CPU. Interestingly, though, they take the same timesteps and end at the same time.

[maximumcats]

The difference is smaller without refluxing on: O(0.1) -> O(1e-5).