Refactor CUB test_device_histogram

* Unify generic and half overloads
* Drop backend selection, since only CUB backend is used
* Refactor transform ops
* Avoid naked new
* Drop unused include
* Remove array extents for function parameters that decay to pointers
* Fix warnings
* Remove CDP temporary storage allocation
* Reorganize dispatch
* Simplify InitializeBinsWithSolution
* Refactor channel params selection
* Replace caching allocator by thrust vectors
* Refactor InitializeBinsWithSolution
* Other stuff