[CUB] DeviceReduce::Min/Max reject gpu_to_gpu determinism, but Reduce with cuda::minimum accepts it

`cub::DeviceReduce::Min` (and related) fails to compile with `gpu_to_gpu` determinism requirement with a static_assert: "gpu_to_gpu determinism is not supported". However, calling `cub::DeviceReduce::Reduce` with `cuda::minimum<>{}` with gpu_to_gpu requirement compiles and runs fine.

Both should compile fine, as gpu_to_gpu is partially supported i.e. well-known cuda operator and integral types.