Reimplement JIT

[hikettei]

Workload

• Prepreq for ShapeTracker (all vm tests are green)
  • Much Simplifier View Backward #155
  • View: Slice+Uprank Merging #156
• Adding more tests for the new JIT!
  • Adding expected scheduling behaviour for the new JIT #157
• 4 jit things to make them green
  • Fix: Scalar mutated by try-insert-node is not a pointer #162
  • Introduce Pointer Adress to caten/aasm, and revisit the semantic of :MOVE #160
    • Fix for _view_backward
    • Remove ADD(a+b+c) #163
  • some schedule stuffs
    • Patch: Constant scalar different across kernels #164
    • axis=1 is broadcasting? batch_size? my impl violates pitmans two bit rules.
    • Transformer batch_size=1 should work, reduction backward also should work.
  • INDEX-COMPONENT+Permute Fusion #166
  • Finish Scheduler and Lowerer Things #170
  • Final: Bring Back Memory Planner #172
  • [todo: fix some view/args type inference]
• Fix for CCL
  • [todo: ]
• New Polyhedral Auto Scheduler
  • [todo: ]
• Bring back the Metal backend
  • [todo: ]
• Fix: Transformer batch_size=1
  • Fix: MergeDims #168

When to ready for merging?

• All CIs become green?
• Added GPT2 Inference tests on CI?
  • batch_size=1 will work?
  • Running on command line tools?
• (Optional) Metal Backend
• (Optional) Tile
• Complete Documentation