[BUG]: extra unnecessary cudaStreamSynchronize

### Is this a duplicate?

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

### Type of Bug

Performance

### Component

Thrust

### Describe the bug

`thrust::transform_reduce(thrust::device, ...` flow is the following:
1. DeviceReduce kernel
2. DeviceReduceSingleTile kernel
3. cudaStreamSynchronize
4. `get_value`: cudaMemcpyAsync + cudaStreamSynchronize

The cudaStreamSynchronize in step 3 is unnecessary when `get_value` is going to perform a stream-ordered operation and synchronize.

This was discovered in TeaLeaf and impacts `std::transform_reduce` performance.

### How to Reproduce

Use the transform reduce algorithm on trivial types on device memory or unified memory.

### Expected behavior

No unnecessary cudaStreamSynchronize.

### Reproduction link

_No response_

### Operating System

_No response_

### nvidia-smi output

_No response_

### NVCC version

_No response_