Optimize memory use on GPUs

[pdziekan]

Some thrust algorithms allocate temporary memory (probably not a complete list):

• sort (_by_key too)
• reduce (_by_key too)
• merge
• scan
• accumulate
• remove_if

We could make them use our own temporary buffers (tmp_device_real_ etc.).
Some direction how this could be done:
https://github.com/thrust/thrust/blob/master/examples/cuda/fallback_allocator.cu or
https://groups.google.com/forum/#!topic/thrust-users/nXQD6j7_xfs