MVP for `cuda::execution::runs_on`

`cuda::execution::runs_on` is intended for use with `cuda::execution::guarantee`, to stipulate that a particular algorithm is expected to execute on the specified GPU or accelerator device. `runs_on` should be a CPO which at the very least exposes call operators for:

- `compute_capability`

This issue can be closed by merging a PR that adds `cuda::execution::runs_on` that allows specifying compute capability and upper bound of number of SMs for use in a single CUB algorithm.