CUB uses THRUST_NS_QUALIFIER::cuda_cub::launcher::triple_chevron for launching all of its kernels. Since all of Thrust kernels will migrate to cub (#26), there's no reason to keep the kernel launcher in Thrust. Unlike remaining tasks in #26, this one is not blocked on VSMem work (#548).

Regarding breaking API, we consider thrust::cuda_cub to be an implementation detail, so strictly speaking there should be no need in maintaining the API in the thrust namespace. Nevertheless, I can see people using thrust::cuda_cub directly, probably because there's no ::detail:: in the name. To be on the safe side, we can add using triple_chevron = cub::detail::triple_chevron.

On the CUB side, we'll need to make sure that triple_chevron header is placed in cccl/cub/cub/detail and the struct itself is in cub::detail:: namespace. Besides that, I'd like to see the test coverage for it. The test should use cdp_launch (example is in /cccl/cub/test/catch2_test_cdp_wrapper.cu) and check that triple_chevron is able to launch kernels from both host and device code. Launching a wrong configuration to validate the return code (example in /cccl/cub/test/catch2_test_debug.cu) would also be helpful.