Add libcudacxx as dependency

[tarang-jain]

There are some places in the codebase where the return type of __device__ functions is queried from host code. One such example is the use of std::result_of in cpp/include/raft/matrix/gather.cuh. The problem is that the return type of __device__ functions might not be correctly fetched from host code using such std headers. Some references to such concerns from the past are:

• Emit diagnostic when attempting to use extended device lambda in contexts that require querying the return type NVIDIA/thrust#1688
• Consider always discarding __device__ lambdas' results NVIDIA/thrust#779

A fix therefore is to use cuda::std::result_of with cuda::proclaim_return_type to correctly obtain the return type of device functions (the corresponding headers are #include <cuda/std/type_traits> and #include <cuda/functional> respectively from libcudacxx), which indicates that this issue is a request to include libcudacxx as a raft dependency. This concern was raised in PR #1445.

[cjnolet]

Just a note- we are planning to add this as a top-level dependency of RAFT soon anyways since we eventually need to migrate our mdspan/mdarray implementations to use those already in libcu++. The main gripe with the changes referenced above is that the dependency was added in the middle of a release in a somewhat unrelated PR. When we do add this dependency, we should do it in intentionally, make an announcement to community, and have it on our roadmap to alleviate surprises downstream.