Enable Double input type for RandomForest and ZipMap

[BowenBao]

• Regardless of input type, attribute types remains Float for floating attributes of RandomForest.
• Regardless of input type, probabilities output type of RandomForest remains Float.
• ZipMap only supports Float input & output type. Update shape_calculator to infer output type from input type.

Signed-off-by: BowenBao bowbao@microsoft.com

[xadupre]

Until ml opset 1, float attributes were the only supported ones (see https://github.com/onnx/onnx/blob/main/docs/Changelog-ml.md#ai.onnx.ml.TreeEnsembleClassifier-1). It changes in ml opset 3 (https://github.com/onnx/onnx/blob/main/docs/Operators-ml.md#ai.onnx.ml.TreeEnsembleClassifier). This was introduced to fix discrepancies due to the float attributes (see https://onnx.ai/sklearn-onnx/auto_tutorial/plot_ebegin_float_double.html).

[BowenBao]

@xadupre Thanks for info. Hence for opset 3, I only need to update the output type of TreeEnsembleClassifier, and type for ZipMap. Attribute type can remain double, unless it's opset <= 2.

ONNXRuntime does not yet support opset 3, so customer needs to specify onnx ml opset 2 for export. Is that achievable through target_opset argument in convert_sklearn? The description says number, for example, 7 for ONNX 1.2, and 8 for ONNX 1.3, which seems like onnx opset but not onnx ml opset.

[xadupre]

The build is fixed in PR #827 (an example in the documentation refers to an onnx model on github and its location has changed).