Make core stepping algorithm asynchronous

[sethrj]

Derived from https://github.com/celeritas-project/docs/issues/17#issuecomment-2985727260 and associated dsicussion

Ideally we could launch an entire step without having to wait on a CUDA return result, so that we can do useful work on the CPU in the meanwhile. There are several aspects of this can be tackled more or less independently, so we can add code to asynchronously...

• [ ] Copy primaries to the GPU (precondition for the first step): #1967
• [ ] Copy hits back to the CPU