We should bullet proof the find_ml_task in EstimatorReport

[glemaitre]

It is not an actual bug but reading again this line:

self._ml_task = _find_ml_task(self._y_test, estimator=self._estimator)

in the initialization stage makes me think that we should double check that we are fine (and maybe more tests). Basically, since we allow y_test to be None sometimes, I want just to be sure that estimator is not in a stage where everything is unifitted and None.

From the top of the head, we should be fine because we should a fitted estimator when y_test is None but we should double check.

[augustebaum]

Related to #1080