replace activemask with ballot function

[DieGoldeneEnte]

This is done to prevent retrieving the wrong thread mask. This is also implemented in hip, but most likely not necessary.

I tested this on an AMD GPU, but didn't do any performance measurements.

fixes #823

[DieGoldeneEnte]

@hipSYCL-ci-bot test please :)

[fodinabor]

Tested on a mixed GPU setup, seems good.
Both HIP and CUDA group_reduce and group_inclusive_scan seem to be in the same performance ballpark as before (according to your sycl-bench fork :))

[illuhad]

Thanks!