Failed loop unrolls

[GMNGeoffrey]

We get some compiler warnings about loop unrolling failing, like this:

/home/gcmn/src/dgl/third_party/HugeCTR/gpu_cache/src/nv_gpu_cache.cu:64:3: warning: loop not unrolled: the optimizer was unable to perform the requested transformation; the transformation might be disabled or specified as part of an unsupported transformation ordering [-Wpass-failed=transform-warning]
   64 |   for (size_t i = lane_idx; i < emb_vec_size_in_float; i += warp_size) {
      |   ^

Based on discussions like NVIDIA/cub#246 and https://gitlab.com/libeigen/eigen/-/issues/1751, I suspect this isn't a big deal and the explicit compiler pragma should actually maybe be removed or the warning silenced. Unclear if that's a change that should go upstream.