[CUCO] Stride tuning for count and retrieve kernels

Similar to https://github.com/NVIDIA/cuCollections/pull/823

The default stride-1 kernel configuration may not deliver the best performance. During the cuco migration, we should invest an effort similar to the `retrieve` optimization to evaluate whether using a larger stride (e.g., 4) can improve performance for `count` and other kernels.

Start by reviewing the cuCollections PR 823 and the current count and retrieve kernel configurations. Benchmark the default stride-1 setup against larger strides such as 4 for count and other relevant kernels; done means performance data supports and identifies any worthwhile tuning.

Written by the indexing model from the issue text.