Return Python float instead of numpy.float64 in sklearn metrics

[lewtun]

This PR converts the return type of all sklearn metrics to be Python float instead of numpy.float64.

The reason behind this is that our Hub evaluation framework relies on converting benchmark-specific metrics to YAML (example) and the numpy.float64 format produces garbage like:

import yaml
from datasets import load_metric

metric = load_metric("accuracy")
score = metric.compute(predictions=[0,1], references=[0,1])
print(yaml.dump(score["accuracy"])) # output below
# !!python/object/apply:numpy.core.multiarray.scalar
# - !!python/object/apply:numpy.dtype
#   args:
#   - f8
#   - false
#   - true
#   state: !!python/tuple
#   - 3
#   - <
#   - null
#   - null
#   - null
#   - -1
#   - -1
#   - 0
# - !!binary |
#   AAAAAAAA8D8=

[lewtun]

I opened an issue on the sklearn repo to understand why numpy.float64 is the default: scikit-learn/scikit-learn#20490

[lhoestq]

Thanks ! :)

[lhoestq]

It could be surprising at first to use tolist() on numpy scalars but it works ^^

[lhoestq]

did the same for Pearsonr here: #2614