Significant discrepancies between `HistGradientBoostingRegressor` scikit and ONNX output

[khoover]

When using skl2onnx for exporting pipelines using HistGradientBoostingRegressor, we are seeing significant discrepancies in the output (double-digit relative error on some inputs) even when we use the CastTransformer to float32-cast inputs to the regressor. For reference, our pipeline looks something like

Pipeline([
    ("preprocessor", ColumnTransformer([("categorical", TargetEncoder(), categorical_features), ("numeric", "passthrough", numeric_features)])),
    ("cast", CastTransformer(dtype=np.float32)),
    ("regressor", HistGradientBoostingRegressor())
])

My suspicion is that the way histogram-based gradient boosting generates buckets and decision thresholds requires the same float64->float32 threshold adjustment that the normal GradientBoostingRegressor gets, even when the inputs are float32.

[xadupre]

I still need to update the converter to support version 5 of the trees which do support double. I did not find the time so far.