[cudauvm] Move conditions to use managed memory

[makortel]

Part of #43. Kind of equivalent to cms-patatrack/cmssw#157.

By default cudaMemAdvise() and cudaMemPrefetchAsync() are enabled, but there are compile-time switches to disable them.

[makortel]

Passes tests, but I want to assess performance impact (on V100) first before merging.

[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

• Managed memory gives 0-4 % smaller throughput than explicit memory
  • Consistent with observations in Use unified memory for conditions cmssw#157
• cudaMemAdvise(cudaMemAdviseSetReadMostly) has a 2-6 % effect until the throughput plateau is reached
  • at the plateau it does not make a difference
• Prefetch has no impact in this test
  • as expected, conditions data is transferred once and then used many many times