Rewrite backend

[jw1912]

Main work is in making the compiler much more serious. This should make it overall easy to write any operation that does something involving a bunch of pointwise/reduction/broadcast/matmul using a simple builder pattern similar to the current GraphBuilder (and get something reasonably well optimised).

Stuff:

• Core generic IR struct that other IRs use internally
• TensorIR
  • Autograd
  • Core ops
    • Pointwise
    • Matmul
    • Broadcast/Reduce
    • Pad/Slice
    • Sparse matmul ops
    • Select ops
  • Optimisation passes
    • Autograd-level passes
    • Generic passes e.g. CSE, DCE, const-folding
    • ScalarConstant stuff
• GPU backend
  • Runtime
    • CUDA bindings
    • ROCm (HIP) bindings
  • FusedPointwise compiling
    • PointwiseIR
    • Compile from a given FusedPointwise
    • Automatically vectorize loads/stores ~hacky
    • Ops
      • Pointwise
      • Broadcast
      • Pad/Slice
      • Select Pad/Slice
      • Sparse matmuls
  • Compiling TensorIR to executable function
    • Matmuls
    • Lower other ops to FusedPointwise individually
    • Fuse adjacent FusedPointwise where possible
      • Check if possible
      • Create fused op
      • Estimate cost
      • Perform highest reward fusion at a time - greedy
• Integration with bullet_lib
  • With mock runtime
  • Sub in actual GPU runtime