`DeviceSegmentedRadixSortKernel` uses `volatile` temporary storage

`DeviceSegmentedRadixSortKernel` uses `volatile` temporary storage:
https://github.com/NVIDIA/cccl/blob/a0a6bfdb5bd7339faa308000db438660df204db4/cub/cub/device/dispatch/kernels/kernel_segmented_radix_sort.cuh#L142-L151

This is at best a code smell, but probably either buggy and inefficient (disabled optimizations). The use of `volatile` should be replaced by proper loads and stores, atomic or regular, as needed and in accordance with the memory model.