Keras metric callback

[Rocketknight1]

Hey, here's the unfinished Keras metric PR, co-authored with @merveenoyan!

Still to do:

1. The outputs from predict/generate are not being concatenated properly, this will be quick to fix! (Edit: Done)
2. I haven't tested prediction with generate at all. generate() is still very slow, but that's a separate PR 😨
3. More broad testing with multiple metric functions.

[merveenoyan]

I feel like it makes sense to test on three notebooks. (could be overkill so you decide)

• Token classification
• Translation
• Masked language modeling
  If you approve I’ll test for these ones. I wonder how it would look like with multiple metrics so in one of those I could test with multiple metrics (prolly token classification one)
  @Rocketknight1

[sgugger]

Thanks a lot for the PR! I think there are a few points to polish but it should be ready to merge soon!

[LysandreJik]

Very nice! You mentioned that this will only work with the latest version of TensorFlow, should there be a check to ensure that users trying to use this for earlier versions are not surprised when it doesn't work for their setup?

[Rocketknight1]

@LysandreJik Absolutely yes, that's getting added!

[sgugger]

You'll need to rebase on master and convert all docstrings to Markdown :-)

[sgugger]

A few last nits. And I think it's still missing the check for the minimum version of TensorFlow?