Model.ConvertToOnnx() fails for Classification pipelines if non-default Features column name is used

[glebuk]

The following code fails to convert to ONNX model:

            var mlContext = new MLContext();
            var data = mlContext.Data.LoadFromEnumerable<Observation>(Data);
            var chain = mlContext.Transforms.Concatenate("FeatureVector", "Price")
                .Append(mlContext.BinaryClassification.Trainers.LogisticRegression(
                    "Label", "FeatureVector"));

            var transformer = chain.Fit(data);
            using (var stream = File.Create("foo.onnx"))
                mlContext.Model.ConvertToOnnx(transformer, data, stream);

On the other hand, if you rename "FeatureVector" to "Features" it works.
This happens for various learners and both binary and multiclass. for both calibrated and non-calibrated. This does NOT happen on regression