[cudauvm] Migrate BeamSpot transfer to managed memory

[makortel]

Part of #43.

[makortel]

Here is a throughput scan on a single V100 (all cases process 400k events / concurrent event, "1 concurrent event" case takes about 9 minutes)

Conclusions

• With advise and prefech (top-right) there is no significant difference wrt. [cudauvm] Move conditions to use managed memory #53
• In the "advise disabled" case (middle-right) [cudauvm] Migrate BeamSpot transfer to managed memory #54 actually improves throughput for small thread count
• In the "prefetch disabled" case (bottom-right) [cudauvm] Migrate BeamSpot transfer to managed memory #54 decreases throughput up to the plateau (<= 10 threads), and increases above that

The throughput degradations (either wrt. explicit memory, or managed memory with advise and prefetch) appear to be clear in the regime where there are free "GPU resources", but become fuzzy when "GPU is fully utilized" (whetever these terms really mean in practice).