Consistent metric outputs

[lhoestq]

To automate the use of metrics, they should return consistent outputs.
In particular I'm working on adding a conversion of metrics to keras metrics.
To achieve this we need two things:

• have each metric return dictionaries of string -> floats since each keras metrics should return one float
• define in the metric info the different fields of the output dictionary

In this PR I'm adding these two features.
I also fixed a few bugs in some metrics

#867 needs to be merged first

[lhoestq]

I keep this PR in stand-by for next week's datasets sprint. If the next release is 2.0.0 then we can include it given that it's breaking for many metrics

[albertvillanova]

Metrics are deprecated in datasets and evaluate should be used instead: https://github.com/huggingface/evaluate