[FEA]: Add support for remote shared memory to barrier<thread_scope_block>

- [ ] Document limitation of `barrier` in peer CTA dsmem

Proposed wording:
```
All member functions of the class `cuda::barrier` have the following additional semantics:

* *Preconditions*: If `*this` is an object declared with the `__shared__` specifier, or a subobject of such, the calling thread must be in the same CUDA thread block as the thread that invoked the constructor of `*this`.
*
```

See https://github.com/NVIDIA/libcudacxx/discussions/325 for discussion.