Fork Optimizations

We should implement the following fork optimizations/analyses

• Fork coalescing
• Fork tiling/untiling (need to figure out how to derive valid tile sizes)
• Fork fusion
  • First doing unrelated forks
  • Then store-load forwarded version through parallel reduce
• Fork fission
  • A version that splits all reduces into separate forks
  • A version that fissions based on an intermediate data, places it into an array, and then uses that array in the following fork
• Fork interchange
• Better forkify (scalar evolution)
• Tensorize (compute einsum notation for each reduce of a fork)