Thrust Babelstream performance on Dot is lacking

Here are results of running [BabelStream](https://github.com/UoB-HPC/BabelStream) with the Thrust and the CUDA backend on my RTX 5090:

Thrust:
```
cmake -Bbuild -H. -DMODEL=thrust -DCMAKE_CUDA_COMPILER=nvcc -DCMAKE_CUDA_ARCHITECTURES=120

Function MB/s Min (sec) Max Average
Copy 1670995.842 0.00032 0.00033 0.00033
Mul 1514830.033 0.00035 0.00037 0.00036
Add 1568846.894 0.00051 0.00052 0.00052
Triad 1576509.679 0.00051 0.00052 0.00051
Dot 1155684.476 0.00046 0.00053 0.00048
```
CUDA:
```
cmake -Bbuild -H. -DMODEL=cuda -DCMAKE_CUDA_COMPILER=nvcc -DCUDA_ARCH=sm_120

Function MB/s Min (sec) Max Average
Copy 1686406.594 0.00032 0.00033 0.00032
Mul 1540921.480 0.00035 0.00036 0.00035
Add 1562139.546 0.00052 0.00053 0.00052
Triad 1577408.291 0.00051 0.00054 0.00051
Dot 1614901.915 0.00033 0.00037 0.00034
```
As we can see, Thrust is severly behind the CUDA implementation for `Dot`, which uses `thrust::inner_product`, which eventually uses `cub::DeviceReduce` with a `transform_iterator(zip_iterator, std::multiplies{})` and `std::plus`. My best guess is that we are not using a good tuning.

This issue can be closed when the performance of `thrust::inner_product` is at SOL again, or at least outperforms the manual implementation in BabelStream.