Reland 4734af313ba2657f7ec8bd33ae5d5fe9249ab62e that is temporarily reverted in https://github.com/intel/intel-xpu-backend-for-triton/pull/5460 by f1e5aadaca84fa42a07d4d64f7981b07c63b0957. CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19287863486/job/55152692885