Implement alternating GPU wait

Implement alternating GPU wait

When both PME and nonbonded tasks are offloaded, instead of waiting in a
blocking call for each task in a predefined order, we poll the GPU
streams and start the reduction of the forces of the task that finishes
first. This allows overlapping one of the reductions with the GPU
compute/transfer of the task arriving second.

Change-Id: I612a0c5cae54bee04c1d587b98b6fc534e766de6