Improve performance of Device Segmented Reduce on large and very large segment sizes

[srinivasyadav18]

The current implementation of device_segmented_reduce only reaches upto 30% in average for segment sizes between 1M and 100M. We could possibliy optimize it by using two phase segmented reduction by using multiple blocks to reduce a single segment.

Bandwidth utilization of current implementation on H100:

Some quick optimization show's promising results:

Leaving with speed ups of