Performance Improvement Possibilities for Reduce_by_key on B200s

The implementation: https://github.com/nanan-nvidia/cccl/tree/reduce-by-key-lookahead

The core idea of the implementation is that the `reduce_by_key` problem can be seen as two problems: (1) getting the unique runs and determine global index (and this is the part where we should use lookahead, like RLE); (2) given the global index and head_flags computed from (1), we just need to reduce the values and this sub problem is now fully parallel. Therefore, we launch 2 kernels, one for each sub problem.

For (2), we notice that the optimal strategy is different for different expected run lengths. Therefore, the implementation is density-adaptive: for each warptile runcount, we change the items_per_thread with corresponding optimizations.

Limitations: only works for 4 byte types as of now, and we need more tempstorage now that we needed to store head flags between (1) and (2).