[FEA]: Investigate performance regression with `cuda::std::transform_exclusive_scan`

### 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.

In #8061 we see performance regressions for large arrays, e.g

```
| I8 | 2^28 | 435.355 us | 0.84% | 460.450 us | 0.71% | 25.095 us | 5.76% | SLOW |
| I64 | 2^24 | 184.862 us | 0.64% | 187.691 us | 1.12% | 2.829 us | 1.53% | SLOW |
| I64 | 2^28 | 2.912 ms | 0.14% | 2.916 ms | 0.21% | 4.395 us | 0.15% | SLOW |
| F64 | 2^28 | 2.872 ms | 0.24% | 2.910 ms | 0.20% | 38.089 us | 1.33% | SLOW |
```

### Describe the solution you'd like

We should investigate where the performance differences come from and see whether we can fix them

### Describe alternatives you've considered

_No response_

### Additional context

_No response_