[FEA]: Explore potential improvements for DeviceMemcpy::Batched

There are a few potential improvements that are worth exploring:

### Tasks
- [ ] Partition work amongst different warps within a thread block. Have one warp perform the decoupled look-back, while the remaining warps partition the buffers by their size (i.e., `tlev`, `wlev`, and `blev`). [Related comment](https://github.com/NVIDIA/cub/pull/359#discussion_r1052178292)
- [ ] Explore whether it's worth to keep around the option `BUFFER_STABLE_PARTITION`. I.e., whether there are cases where stable partitioning may turn out to be more performant. [Related comment](https://github.com/NVIDIA/cub/pull/359#discussion_r1052028739).
- [ ] Make use of [builtin-assume-aligned](https://docs.nvidia.com/cuda/cuda-c-programming-guide/#builtin-assume-aligned) to simplify `VectorizedCopy` and see how it compares to the current implementation.