Improve GPU Load-Balancing / job pools

[olepoeschl]

Currently we have a static approach for balancing gpu block load and balancing load between multiple gpu's. This works fine for smaller board sizes.

However, with increasing board size, the computation wastes a lot of time in the end because the gpu is not 100% occupied any more or respectively some gpu's are already done while one is still working.

Solution: We should find a dynamic approach to balance the load between gpu blocks and also balance the load between multiple gpu's. It would be optimal if the approach addresses both issues at once.

Proposal: Have the whole job pool in CUDA / OpenCL pinned memory that each device can access. Keep in mind that all threads in a gpu block require to have equal jkl's.