improve lumi counters
Now a block is scheduled per event instead of having a fixed number of blocks that process all events.
In order to get deterministic results for the counters, a intra-block copy of them in created in shared memory, using unsigned long type, and the results are added back to global memory at the end. This, in addition to MooreOnline!395, makes the counters (and their tests) stable across different runs
Time spent on these algorithms decrease by an order of magnitude (see current.txt and branch.txt files) and overall throughput increases ~13%.