New Accelerators: NATIVE, Metal, and placeholders for CUDA, LLVM, WebGPU

[hikettei]

• CI: Test w/ Lisp Codegen (TODO: Rename LISP_JIT -> NATIVE)
• Move to native.lisp (add Readme)
• Nodes inside the EXPR may also want type relay (for optimizing NATIVE)
• Add: BACKEND=NATIVE
  • Threefry2x32 is failing due to wrap-around (TODO: Replace them with MOD)
  • TODO: Support parallel in default (if the :FOR is enough simple though...)
  • Make the CI green
• BACKEND=METAL
  • kernel worked
  • threefry2x32: needs to transfer the buffer: CPU -> Metal
  • Tested in CI
  • Optimize global/local size
  • Use MTLCompiler directly for the compilation speed
• TODO: BACKEND=CUDA
  • Add: Renderer
• TODO: BACKEND=LLVM
  • Add: Renderer
• TODO: BACKEND=WGPU
  • Add: Renderer
  • Is there any decent wrapper for webgpu?
• CUDA: Looking into Petalisp? or using cl-cuda? inconveniently i dont have cuda computers so i cannot test locally :(