Update examples to show how to deal with extra validation copies

[muellerzr]

Update examples to show how to truncate the validation set for metrics

What does this add?

Based off this issue this PR updates all examples to show how to get rid of extra samples that get added when performing distributed training on the validation set.

Testing on a multigpu system will happen tommorow, but @sgugger pretty sure the way I have it setup ensures that this only runs when we have distributed systems, and that's where this problem arises?

Who is it for?

Should close #287

Why is it needed?

It's unclear from the scripts how to alleviate this behavior, and it's not documented anywhere. So, with this PR it now is

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[muellerzr]

Note: This is just an initial to make sure the format and whatnot looks right and then all the other examples will follow suite :)

[sgugger]

Thanks for the PR! I'd just put this in a specific feature example instead of the base one.

[sgugger]

Thanks! Not sure what your question is about tests. To test this we would need to know in advance the exact value of the metric and make sure we get that again but it's very finnicky since the metric computed in the base script is roughly the same.