Add more metrics support in skelarn backend

[DrAnaximandre]

So far only the mean absolute error is supported. We could increase that number of supported metrics in a future release.

• A first step could be to include other sklearn.metrics, that should be quite easy.
• A second step could be to test the serialization of custom objects.

[DrAnaximandre]

PR to come soon. Local modifications:

• [x] switch the default metric to the score attribute of the sklearn model
• [x] support metrics that are available in sklearn.metrics
• [x] change tests so that they test 0, 1 and 2 additional metrics
• [x] local validation

---

• The major modification in terms of behavior for the end-user is that the model.full_res.metrics now has a field score, that is related to the score attribute of the sklearn model (that is to say: the user should know what the behavior of its model is).
• The user can now specify a list of metrics in the fit, for instance: expe.fit([data], [data_val], metrics=['accuracy_score', 'mean_squared_error']) They will be computed on the top of the score thus lead to a number of additional predictions with the model depending on the number of elements in data and data_val (more specifically: the prediction is done regardless if there are additional metrics and used for all the additional metrics). That leads to 2 optimization points:
  • it would be easy to add a check and do not do the additional prediction if there are additional metrics.
  • however getting the predictions from the score function of the sklearn model is not possible. A possibility would be to hard-code the default metric for all supported sklearn models if we really want to avoid that unnecessary prediction.