Full-sweep performance evaluation for non-deterministic `DeviceBatchedTopK`

As part of https://github.com/NVIDIA/cccl/issues/9253 we want to evaluate the performance across a full sweep of segment sizes, batch size, and k, for different input distributions.

## Tasks

- [ ] Evaluate options to compare performance of RAFT or XLA versus CUB
- [ ] Get preliminary performance results from RAFT or XLA

A similar evaluation was conducted by the XLA team to identify qualifying implementations:

The issue can be closed with performance results similar to the one above comparing against RAFT or the current state of XLA top-k routing.

No file or test entry point is named. Start with the linked issue 9253 and evaluate RAFT or XLA against CUB across segment sizes, batch sizes, k values, and input distributions; done means performance results comparable to the referenced XLA evaluation.

Written by the indexing model from the issue text.