Be more tolerant in check_array for CalibratedClassifierCV

For our package http://github.com/metric-learn/metric-learn, the function CalibratedClassifierCV is very convenient for Weakly Supervised Learners, as it can make PairsClassifier estimators return a probability for a pair of points to be labeled as similar or dissimilar, when those return a decision function.

However, we currently cannot use it because our inputs can be 3D (example: pairs=[[[2.3, 5.4], [4.4, 5.6]], [[7.5, 1.2], [4.4, 5.6]]]), and CalibratedClassifierCV uses check_array with default parameters that does not allow 3D inputs.

However, other meta-estimators like GridSearchCV do not call check_array, so we can use them easily in metric-learn.

Is the check_array in CalibratedClassifierCV really useful or could we do without it ? If we could, I’d be happy to provide a PR to do so

[connorbrinton]

I think this issue is also related to #8710, which has to do with CalibratedClassifierCV’s input validation in a pipeline context

[bellet]

So would the addition of allow_nd=True in the check_array be the way to go then?