[EPIC]: `mdspan`-based algorithms

### Is this a duplicate?

- [x] I confirmed there appear to be no [duplicate issues](https://github.com/NVIDIA/cccl/issues) for this request and that I agree to the [Code of Conduct](CODE_OF_CONDUCT.md)

### Area

libcu++

### Is your feature request related to a problem? Please describe.

1. `mdspan` represents the C++ abstraction for multidimensional data. We currently support only a small subset of available features
2. The `llm.c` version presented to the [CUDA MODE](https://github.com/cuda-mode) group can be greatly simplified and improved by supporting additional features related to `mdspan`

### Describe the solution you'd like

- Provide a `cuda::` `streaming accessor` instead of manually instantiate it
- Allow `mdspan` to support device-only/host-only code for custom accessors and layouts
- Provide `for_each_in_extent` to directly linearization/de-linearization multidimensional indices
- Provide `mdarray` (owning version of `mdspan`)
- Provide a vocabulary type to declare host or device `mdspan/mdarray`

### Describe alternatives you've considered

_No response_

### Additional context

_No response_