[FEA]: Provide `cuda::std::constant_wrapper`

`std::constant_wrapper` is a fancy wrapper for constant values introduced in C++26 that tries to act like it's the original type. It could be useful in cases you want to have 1 generic kernel to which you'd like to either a constant or a value.

This is a generally useful utility. I would also suppose that more `std::` APIs would take the constant wrapper as a parameter in the future. It's already part of `std::function_ref` design.

This issue can be closed by fully implementing the `cuda::std::constant_wrapper` in libcu++. The implementation is not too complicated, but for our project, there are some challenges we need to deal with:
1. Compiler bugs. The implementation uses features that were not that popular, thus less used, thus less explored, thus less bugs were found.
2. Passing the type from host to device, making sure we don't access host memory from device.

#### Blocking nvbugs
- 6139224, no workaround found yet, clang-cuda is fine with the code
- 6207656, needs to be resolved, clang-cuda is fine with the code

#### Other reported nvbugs
- 6183205, workaround: using SFINAE instead of concepts
- 6183211, workaround: using `decltype(constant_wrapper<__cw_fixed_value(X op= Y)>{})`