catch tensor.numel() == 0 in nan detector (pytorch#140741)

Context: we are trying to pass an empty tensor through the system now (sometimes;... its an edge case); and it seems to cause all_reduce to seg fault, which is unexpected to me

Deep Shah and Pavan identified the issue, I'm just pushing for a fix :)

Test Plan: idk what i'm doing here, someone help

Reviewed By: shuqiangzhang

Differential Revision: D65956095

Pull Request resolved: pytorch#140741
Approved by: https://github.com/shuqiangzhang

torch/csrc/distributed/c10d/NanCheck.cu

@@ -233,6 +233,10 @@ void checkForNan(const at::Tensor& tensor, at::cuda::CUDAStream& stream) {
   const size_t numThreadsPerBlock =
       std::min<size_t>(maxNumThreadsPerBlock, tensor.numel());
 
+  if (!(numThreadsPerBlock > 0)) {
+    return;
+  }
+
   const size_t numBlocks = std::min<size_t>(
       maxNumBlocks,
       (tensor.numel() + numThreadsPerBlock - 1) / numThreadsPerBlock);