Add bitonic sorting kernels

[upsj]

This PR adds a three-stage bitonic sorting implementation for GPUs:

• In the first stage, a power-of-two number of elements are being sorted locally within each thread
• In the second stage, these local elements are being sorted within a warp/wavefront using butterfly shuffles
• In the third stage, these elements from a power-of-two number of warps/wavefronts are being sorted using shared memory. Here I took additional care to avoid shared memory collisions.

Note that in a bitonic sorting network, we first recursively sort two halves of the input, with the second half reversed, and then merge the result by first merging corresponding elements from the two blocks and then recursively merging the elements inside the two halves.

TODO:

• Add documentation describing how to invoke the helper functions

[thoasm]

Request for a some static_asserts to restrict template parameter range (therefore, making it easier to find potential errors in the code).

[thoasm]

LGTM!

[pratikvn]

A reference (paper) to where the bitonic sort algorithm has been first proposed would be nice. Anywhere in the class documentation would be good. Otherwise, LGTM.