[FEA]: Provide way API for forcing mapping to produce 64-bit wide ranks/counts

Currently, I use `unsigned` for storing counts/ranks everywhere. However, this may not be sufficient when grouping threads in grid. We should provide a way to let the mapping know to use 64-bit integers. Something like:
```cpp
cuda::my_mapping mapping{cuda::use_64bit}
```
We should also either `uint32_t` or `uint64_t` types instead of plain `unsigned`.