Improvements over bce4

[akamiru]

• Completly vectorized: 16 "butterflies" each iteration
• Added fast ANS encoder and decoder
  • Reduced compression compared to the adaptive
  • A LOT faster able to keep up with the new level of performance
• Changed iteration layout
  • Handles 8 queues interleaved instead of one after another
    • Greatly improved cache locality: -75% cache misses
  • Allows caching of rank queries and part of the queues
    • Greatly reduces the number of gathers and scatters: -50% µops
    • Reduces memory usage of the queues by 57%
• Improved rank update
  • Clears and sets bits base on butterfly variables rather then lzcnt/tzcnt moving bits
  • Updates the prefix sum (base rank) only iff necessary rathe rthan recalculating it every time
  • Never needs to gather the prefix sum (base rank)
  • Logarithmic conflict resultion reduces writes to memory