Test MPI/CUDA polling in tasks instead of integrating with scheduler

[msimberg]

We should test if running the MPI/CUDA polling in standalone tasks works as well as integrating the polling in the scheduler. This would make the polling independent of the scheduler that it's running on and possibly make integration with other libraries easier. In addition it might make shutdown a bit simpler since we don't have to separately keep track of task counts, MPI request counts, and CUDA event counts; everything would fall under the task count. This might also make the polling slower. This also similar to what is proposed in ALPI.