Launcher should not collect all the workers' logs

I do not know why the launcher will collect all the logs. Just copy the **rank0's** looks more useful. Because i want to print some information in woker1 for debug. But it will be robbed by the laucher.
```
# Horovod: print output only on first rank.
if hvd.rank() == 0:
....
print('\nTest set: Average loss: {:.4f}, Accuracy: {:.2f}%\n'.format(
test_loss, 100. * test_accuracy))
# Horovod: print test only on other rank.
if hvd.rank() != 0:
print('\ntest!!!\n')
```
```
$ kubectl logs pytorch-mnist-tb-launcher-cnk95
....
Train Epoch: 6 [29440/30000 (98%)] Loss: 0.211814
Train Epoch: 6 [29440/30000 (98%)] Loss: 0.098771

test!!!

```
It make me confused

The issue names no source files or tests. Start by tracing how the MPI launcher exposes worker output through kubectl logs, then reproduce the rank0 and non-rank0 prints shown in the report; done means the launcher behavior for worker logs is defined and matches the intended output without hiding needed debug information.

Written by the indexing model from the issue text.