Investigate the performance impact of CUDA managed memory

[makortel]

I'm thinking to make a full version of cuda program that uses the CUDA managed memory (kind of follow-up to cms-patatrack/cmssw#85). The question of the performance impact seems to affect both the portability technology exploration and the evolution of the CUDA products in EDM, so maybe it is time to make a proper evaluation.

Plan

• Clone cuda to cudauvm done in Add cudadev and cudauvm programs, update README #44
• Clone CachingHostAllocator to CachingManagedAllocator, create cms::managed::unique_ptr done in [cudauvm] Introduce CachingManagedAllocator and managed_unique_ptr #49
• Migrate conditions done in [cudauvm] Move conditions to use managed memory #53
• Migrate beamspot transfer done in [cudauvm] Migrate BeamSpot transfer to managed memory #54
• Migrate raw-to-cluster done in [cudauvm] Migrate clustering to use managed memory #123
• Migrate rechits done in [cudauvm] Migrate rechit to use managed memory #162
• Migrate CA done in [cudauvm] Migrate tracking and vertexing to unified memory #182
• Migrate vertexing done in [cudauvm] Migrate tracking and vertexing to unified memory #182
• Evaluate performance, at least on T4 and V100

I'm thinking to have compile-time switches to

• enable/disable cudaMemPrefetchAsync()
• use device/managed memory for (temporary) device-only data structures

[makortel]

The evaluation was reported in vCHEP 2021