Add generic reduction functions and separate reductions/warp_primitives

[akifcorduk]

This PR adds bunch of new device reduction functions such as:

• Generic device reductions that takes reduction operator as argument.
• Ranked reductions to return the index/rank of the reduced value.
• Weighted random reduction to have probabilistic reduction using conditional probability.
• Binary reduction to reduce binary values more efficiently.

There are tests implemented for all device reduction operations.

This PR also separates warp primitives to the warp_primitives.cuh.
All reduction functions are moved to reduction.cuh

[ahendriksen]

Thank you @akifcorduk: it's a good idea to split up cuda_utils and add tests. I have added some comments.

[tfeher]

Thanks @akifcorduk for this work! Most of it looks good, but I have a few questions below.

[tfeher]

Thanks Akif for the updates, a few additional comments.

[tfeher]

Thanks @akifcorduk for addressing the issues. Apart from one issue with the right home for the new random primitives it looks good to me, therefore pre-approving.

[cjnolet]

/merge