Optimized mask clustering

• It now finds 0.019493% more clusters (down from 0.07%)
• The algorithm should be a tad faster
• Added support for CMAKE_BUILD_TYPE option. Available options:
  • RelWithDebInfo (default)
  • Release
  • Debug
• EstimateInputSize logic changed for adding candidates. Using masks now.
• Removed sp_size from the GPU (it was unused).
• Added constant candidate_ks for finding out active pixel numbers in a four-bit number (EstimateInputSize optimization).
• Prefix Sum has been optimized, following the strategy of Merrill’s 2–level upsweep/downsweep.
• When profiling Clustering alone, or the whole application, performance rate is now more stable and less picky about synchronization shenanigans.
• A Handler class has been created, holding the stream, blocks and threads attributes. Any Handler should inherit from it.
• Consolidate tracks is on by default now.
• Found a good configuration for EstimateInputSize call. 70 kHz on 1080 Ti.