[Inductor] simplify CPP backend Tile2D code and support non-contiguous load/store (#97626)
Remove `CppTile2DTailKernel` and `CppTile2DKernelChecker` and reuse `CppVecKernel` and `CppVecKernelChecker` for them. Add vectorization with fallback for load/store in CppVecKernel for the non-contiguous load/store needed by `CppTile2DTailKernel`.
This PR also adds a functional support for transposed copy of bfloat16 data types. Better performance requires vectorized intrinsics implemented for at::vec::transpose_mxn. cc @soumith @voznesenskym @penguinwu @anijain2305 @EikanWang @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97626
Approved by: https://github.com/jansel