AOT for the LLVM backend

[k-ye]

• [x] Make CPU backend work
  • [x] Support arg-less kernels
  • [x] Support scalar + Ndarray args
  • [x] Split LlvmProgramImpl into compile-time & runtime parts to form Runtime API
• [x] Make CUDA backend work
  • [x] Support arg-less kernels
  • [x] Support scalar + Ndarray args
• [ ] Provide public interfaces
  • [x] Interface with Compute Graph
  • [ ] Interface with C-API
• [ ] Improvements & Demo
  • [ ] Dense/Pointer SNodes Demo
  • [ ] Support more flexible snode_id, snode_tree_id (determined at runtime)
  • [ ] Implement SNodeTreeType: https://github.com/taichi-dev/taichi/blob/master/docs/rfcs/20220413-aot-for-all-snode.md
  • [ ] Support memory deallocation for CPU backend. Merge Device::allocate_memory_runtime() with Device::allocate_memory()
  • [ ] Verify SNode GC works for Field AOT.
  • [ ] Serialize/deserialize Ndarray/SNodeTree buffers: https://github.com/taichi-dev/taichi/issues/5428