The core registration algorithm of small_gicp is re-written from scratch to improve the processing speed, and I suppose small_gicp might be faster than nano_gicp, which is based on the previous implementation fast_gicp.
DLO/DLIO are scan-to-model matching odometry estimation algorithms while small_gicp is a more primitive scan matching algorithm. So, in terms of the accuracy, it is apples-to-oranges comparison. But, I suppose there is no accuracy differences in the context of scan-to-scan matching.
reference link: [nano_gicp]