[FEA]: Widen CUB benchmarks to cover extended floating-point types

### 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

CUB

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

CUB currently has limited benchmark coverage for extended floating-point data types such as `__half` , `__nv_bfloat16`, FP8, FP4. These types are essential in LLMs, or more in general, in AI workloads. However, most CUB benchmarks still focus primarily on classical C++ integer and floating-point types.

Performance can differ significantly from C++ types due to conversion operations, vectorization, data movement, accumulation, and GPU-specific instructions.

CUB currently only covers `__half` , `__nv_bfloat16` for `DeviceTransform` + `WarpReduce::Sum/Min`.

Benchmarking such types would allow to keep track of regressions and would set baseline for optimization opportunities.

### Describe the solution you'd like

The issue can be closed by adding the benchmarks for:

- `__half` , `__nv_bfloat16`. (potentially FP8, FP4 types but out of scope.)
- Routines: Device/Block/Warp Reduce and Scan at least. To also strongly consider Sort

In order of relevance:

1. `{Device/Block/Warp}Reduce::Sum/Min`
2. `{Device/Block/Warp}Scan::Sum`
3. `{Device/Block}RadixSort` (partially covered by `int16_t`)
4. [optional] Segmented algorithms of the Reduce/Scan/Sort should already benefit from Block/Warp-level improvements of the previous points

Start by locating the existing CUB benchmarks for DeviceTransform and WarpReduce::Sum/Min, then inspect how benchmark coverage is organized for other types. Add coverage for __half and __nv_bfloat16 across the requested Reduce and Scan routines, with RadixSort as a lower priority, and verify the benchmarks build and provide regression baselines.

Written by the indexing model from the issue text.