[clang] [CUDA] Support calling `consteval` function between different target. (#158688)
In CUDA, calling `consteval` functions cross excution space is allowed.
So the function with `consteval` attribute need be treated as a
`__host__ __device__` function.