can sparse all-reduce keep efficiency with large number of gpu workers？

in my opinion, when the gpu cluster scaled up to several hundred workers, high sparsification ratios still generate significant communication overheads, which even worst than DenseAllReduce.

No files, tests, or entry points are named. First locate the sparse all-reduce and DenseAllReduce implementations and any GPU-worker benchmarks, then compare communication overhead as the worker count reaches several hundred and document whether sparse all-reduce remains more efficient.

Written by the indexing model from the issue text.