Workaround for bug in DistributedDataParallel (#46186)
Summary:
Fix the DistributedDataParallelSingleProcessTest to work around a limitation in DistributedDataParallel where the batch_size needs to evenly divide by the number of GPUs used
See https://github.com/pytorch/pytorch/issues/46175
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46186
Reviewed By: bdhirsh
Differential Revision: D24264664
Pulled By: mrshenli
fbshipit-source-id: 6cfd6d29e97f3e3420391d03b7f1a8ad49d75f48