Multi-CTA `cub::DeviceBatchedTopK`: Carve out and add the `BlockPartition` building block

The `BlockPartition` building block maps to two scenarios, as outlined in https://github.com/NVIDIA-dev/cccl_private/issues/659:
1. the `buffered` scenario where "selected" items are written to the user-provided output iterator, candidate items are written to a scratch "candidate" buffer and
2. the `last_filter` scenario with selected items being written front-to-back and the remaining candidates (tied items across the k-th boundary) being capped to not exceed the remaining "free" slots (e.g., growing back-to-front).

The issue can be closed once the PR that adds the `BlockPartition` building block has been merged.