Fix bug where ZeRO2 never uses the reduce method. (#4946)
On this PR https://github.com/microsoft/DeepSpeed/pull/4695, the
gradient synchronization operation is moved to the `allreduce_bucket`
method, but on this method, rank is set to None, and it will never use
the reduce method even if `use_multi_rank_bucket_allreduce` is set to
False.
Co-authored-by: jializheng <jializheng@huawei.com>
