Faster compactions for up to 1024 elements

[kstppd]

• This PR tries to resolve issue Feature request: splitvector methods which run only a single kernel could work without stream syncs #41 and issue Naming request: cudaMemPool to gpuMemPool or, even rather, splitMemPool #46.
• It introduces new block compaction kernels that show performance improvements for compactions of up to 1024 elements ie up to a single block of 1024 threads. Compactions utilize register communication and warp reductions.
• Tested on both NVIDIA and AMD hardware.
• Cuda_mempool is now renamed to SplitStackArena.