[Feature] New Backends: Metal and Lisp

[hikettei]

• Metal Codegen Backend
  • "It just works" is fine. we are not in the stage that we consider the performance.
  • TODO: BEAM Search for determining global_size/local_size
• Lisp JIT
• Some renderer simplification
• Introduce Partial Schedule to eliminate (c0 == 0) and gemm-test
  • Introduce partial schedule and make schedule_whole_components working again
  • investigate why atomic wont work
• 1 Kernel 1 Reduce Rule?

Changes

• Symbolicが綺麗なコード吐けるようになったらstaticでunrollとかreduce accumとかやる
• Fuse Conv < 2 kernels
  • NN Ops Test with accuracy
• Reduce Accum
• Simplifier: Fold Scalars
• If (c0==0)をいい加減削除したい，SymbolicとStaticのコンパイル結果は一致するべき eliminate guards
• threefry2x32 < two kernels
  • fused but wont work
• dynamic shaped kernel compilation (which is the equivalent to a static shape system)
• fix: matmul permute y
• float accumlator
• float4 unrolling
• RaW/WaW depsがうまく読めてない箇所がいくつかある: (SERIALIZE=1でpassed)
  • !rand
    • !rand compilation requires 30s on M3 Pro due to the compilcated and largest kernel
  • view-backward (failing caseも動作した...)
  • etc

[hikettei]

currently ci fails due to dissect?